Rank every tool on the list by what company data goes into it and how many people use it, then deal first with the few receiving regulated or confidential data such as customer records, health or card data, source code, credentials and deal documents. Make one call per tool: sanction it on a business plan with SSO and admin controls, move its users to the AI tool you already approved, replace it with an approved tool that does the same job, or block it. Then decide whether data already sent needs action, which means reading the vendor’s retention and training terms, asking for deletion and assessing whether it counts as a breach. Give staff a no-blame window to self-report before anything is blocked, and open a fast request path so the next tool arrives as a request. Network logs show which AI sites were visited, so the gap is what was pasted or uploaded and from which account. An endpoint agent on the laptop closes it: on managed Windows and macOS, Jazz sees GenAI prompts, pastes and uploads at the operating system level in any browser, desktop app or command-line tool, without a browser extension, and tells a personal ChatGPT or Claude account from the company workspace, so triage starts from evidence.
Updated October 23, 2026.
Which tools can find and control shadow AI, and what each one sees
What separates these approaches is whether they see the data going in and the account receiving it.
| Approach | What it sees | What it misses |
|---|---|---|
| DNS, proxy or CASB logs and domain blocks | Visits and volumes per AI domain on traffic that crosses your network | Personal and corporate accounts on the same domain, laptops off the network, phones, and what was pasted |
| Vendor tenant headers through a TLS-inspecting proxy | Lets only your ChatGPT Enterprise workspace or Google domain load on managed networks | Traffic that never crosses the proxy, and desktop apps that pin certificates unless device management pushes exceptions |
| Expense reports and OAuth grant reviews | Paid personal subscriptions and AI apps holding access to corporate mail and files | Free tools, and apps opened with a personal Gmail account |
| Browser controls such as Microsoft Purview browser data security | Listed consumer AI apps inside Edge for Business on Intune-managed Windows devices | Desktop apps, terminals, and devices outside Intune-managed Windows |
| Endpoint agent (Jazz) | GenAI prompts, copy and paste, file uploads and screen sharing at the operating-system level in browsers, desktop apps and command-line tools, with the account marked personal or corporate | Personal phones, unmanaged computers and AI processing inside a SaaS vendor’s cloud |
| Self-report window and request path | Which tools people use and why, in their own words | Anything people do not disclose |
For a company that already licenses ChatGPT or Claude but finds personal accounts of the same tools on its list, Jazz fits because its agent sees the paste on the laptop and treats the company workspace and a personal login as different destinations, wherever the laptop is connected. SaaS-embedded AI and phones still need admin-console settings and identity controls.
Why shadow AI is important to address now
A list of 40 or 140 tools is normal. When the Cloud Security Alliance published its Jazz-funded DLP in the AI Era survey on 13 October 2026, Jazz co-founder and CEO Ido Livneh said that across the customer environments Jazz observes, the average organization uses nearly 170 AI tools and fewer than 10 are sanctioned. Gartner’s survey of 302 cybersecurity leaders found that 69% of organizations suspect or have evidence that employees use prohibited public GenAI (Gartner, 19 November 2025).
The cost shows up in breach data. IBM’s Cost of a Data Breach Report 2026, published in July 2026 and based on 602 breached organizations, found that security incidents involving shadow AI more than doubled to 43%, from 20% a year earlier. Those breaches averaged USD 5.39 million, data loss or compromise followed in 49% of them, and in about one in five the organization paid a regulatory fine. In the same report, 68% of breached organizations lacked AI governance to manage AI or detect shadow AI (ComplexDiscovery analysis of the IBM report, 1 August 2026).
Visibility is the weak point. In the CSA survey of 433 IT and security professionals, 48% ranked GenAI as their top exfiltration risk, yet only 29% could produce a routine 30-day report showing sensitive data sent to GenAI tools. A domain report shows that a tool was visited, and says nothing about whether someone pasted a customer list into it.
How to identify and mitigate shadow AI risks: sort the list by data and users
The tool count is the wrong measure. Ten people pasting public marketing copy into a slide generator are a smaller problem than one finance analyst uploading an unreleased earnings model to a personal chatbot account. Give every tool on the list four attributes before anyone argues about blocking.
- Data going in. Put each tool in one of three tiers. Tier 1 is regulated or confidential data: personal data about customers, patients or employees, card data, source code, credentials and API keys, unreleased financials, M&A and legal documents. Tier 2 is internal business data such as drafts, internal reports and meeting content. Tier 3 is no company data at all.
- Users. Count distinct people, and note which teams. Fifty engineers on one coding assistant usually means a sanctioning decision, while a single user on an obscure PDF tool can often be settled with one conversation.
- Account type. Record whether people use a personal account, a company account or no account. The same service can appear twice on your list.
- Access mode. Pasted text sits at the low end. File uploads come next, then standing OAuth access to mail, calendars or drives, then agents that act on their own.
Deal with tier 1 tools this week, ordered by users and access mode. Tier 2 tools go through the sanction or replace review this month. Tier 3 tools go to the request path and the acceptable use policy.
Different discovery sources fill different columns. DNS, proxy and CASB logs show visits and volumes. Expense reports show paid personal subscriptions. OAuth grant reports show which apps hold access to corporate mail and files. Only an endpoint or a user’s own report shows what was actually pasted or uploaded, and from which account. How to detect shadow AI in your company compares these sources. Start there if you do not have the list yet.
A first-week and first-month plan
First week
- Merge every source into one inventory with one row per tool and account type, and name one owner for it in security or IT.
- Fill in the four attributes above and mark the tier 1 tools. Where you cannot tell what data goes in, ask the heaviest users directly.
- Send the staff message and open the self-report window described below, before any block goes live.
- Make the per-tool call for each tier 1 tool. Block immediately only where the tool has no business plan, receives regulated data and has an approved alternative ready, or where it is a known data-harvesting extension.
- Run the data-already-sent check for every tier 1 tool. Rotate any credentials or API keys you know were pasted.
- Publish the approved tool list and a short request form.
First month
- Run sanction reviews for the popular tier 2 tools with legal and privacy, and buy seats where a review passes.
- Move personal-account users of licensed services into the company workspace, using the vendor’s domain-claim features where your plan has them.
- Review the AI features inside SaaS apps you already pay for, the OAuth grants held by AI apps, and the browser extensions installed on managed browsers.
- Close the self-report window, fold every disclosed tool into the inventory and follow up individually on tier 1 disclosures.
- Update the AI acceptable use policy to match the decisions you made.
- Re-run discovery and compare it with the first list. New entries show whether the request path is working.
Decision table: tool type and data going in
| Tool type | No company data | Internal business data | Regulated or confidential data |
|---|---|---|---|
| Chatbot you already license (ChatGPT, Claude, Gemini, Copilot) used on personal accounts | Move users to the company workspace | Move users; claim personal accounts on your domain where the plan allows | Move users now, stop personal-account use for this data, assess what was sent |
| Chatbot with a business tier you do not license | Leave to the request path | Sanction on a business plan if usage is wide and review passes; otherwise replace | Replace with an approved tool, or sanction only after legal and privacy review; stop personal use meanwhile |
| Single-purpose web tool with no business plan (PDF summarizer, slide maker, transcription site) | Allow or replace | Replace with an approved tool | Block and assess what was sent |
| AI browser extension | Allow only reviewed extensions | Allow only reviewed extensions | Remove; treat conversations as exposed if the extension is a known harvester |
| AI notetaker joining meetings | Sanction one tool with admin controls | Sanction one tool; keep others out of company meetings | Keep all notetakers out of meetings covering this data until reviewed |
| Coding assistant or command-line agent | Sanction the enterprise tier | Sanction with login restricted to your organization | Sanction only with organization-enforced login and no-training terms; rotate exposed secrets |
| AI feature inside a SaaS app you pay for | Leave on | Check the admin setting and the vendor’s AI terms | Switch off for the affected groups until reviewed |
| Agent or app with OAuth access to mail, drive or calendar | Restrict OAuth to reviewed apps | Revoke and re-approve through admin review | Revoke the grant now and review what it accessed |
Sanction, move, replace or block: how to make each call
Sanction a tool when many people use it for real work and the vendor passes your review. Sanctioning means a business or enterprise plan, which is where the data terms change. OpenAI does not train on ChatGPT Business, Enterprise, Edu or API data by default, lets workspace admins control retention, and signs a data processing addendum (OpenAI enterprise privacy, updated 8 January 2026). Anthropic’s commercial products are not used for training by default (Anthropic, 18 August 2026). Google says content in qualifying Workspace editions is not human reviewed or used for model training outside your domain without permission (Google Workspace, 14 August 2026). Add SSO, set retention and record the decision. How to run an AI risk assessment covers per-tool vetting.
Move users when they are on a personal account of a service you already license. These users already have a company account waiting, which makes this the quickest group to fix. Claude Enterprise admins can see every personal Free, Pro and Max account on a verified company domain and migrate them into the company workspace within a 30-day window; Team plans can only block new personal accounts (Anthropic, 2 September 2026). OpenAI states that ChatGPT Business admins cannot require a personal workspace to be merged or deleted (OpenAI Help Center, October 2026). Accounts opened with a personal email stay out of reach of all of these features.
Replace a tool when the job is legitimate but the tool is not. A one-person PDF summarizer with no business tier rarely justifies a contract, and the approved assistant can usually do the same work. Tell the user what to use instead in the same message that tells them to stop.
Block when the risk is clear and the alternative is ready: tools with no business tier receiving regulated data, extensions caught harvesting chats, and personal accounts receiving data that must stay under company terms. Vandy Hamidi, CISO of accounting firm BPM, describes the order of work as understanding why people use a tool, checking whether an approved tool already meets the need, and then redirecting them (CSO Online, March 2026). A block with no alternative tends to send the work to a phone.
What to do about data already sent to an unapproved AI tool
A block only affects future uploads, so data already in a vendor’s systems needs its own decision. For each tier 1 tool, work through three questions.
What went in, and from which account? You need the data type, the rough volume, the dates and whether the account was personal or corporate. Without that, every later decision is a guess.
What do the vendor’s terms do with it? The answer depends on the plan, and personal plans are the ones on your list.
| Service and plan | Used for model training? | Retention after deletion |
|---|---|---|
| ChatGPT Free, Plus, Pro | May be used unless the user opts out; feedback can still send a whole conversation to training (OpenAI, October 2026) | Permanently deleted within 30 days, with security and legal exceptions (OpenAI, September 2026) |
| ChatGPT Business, Enterprise, API | Not by default | Admin-set retention; deleted conversations removed within 30 days unless law requires otherwise |
| Claude Free, Pro, Max | Only if the user allows it; opted-in data kept up to five years (Anthropic, 28 August 2025) | Deleted from back-end storage within 30 days; data stays in models already trained (Anthropic, 1 July 2026) |
| Gemini apps, personal account | Used to improve Google services, with some human review (Google, 19 October 2026) | Auto-delete defaults to 18 months; chats seen by human reviewers are kept up to three years even after deletion |
Can it be deleted? For a personal account, only the account holder controls deletion in practice. Ask the employee to delete the relevant chats and files, and record that they did and when. Where the vendor offers domain claiming on your plan, migrating the account brings its history under company control. Deletion does not pull data back out of a model that has already been trained on it, and reviewed Gemini chats outlive deletion, so treat deletion as reducing exposure rather than reversing it.
Is it a breach? If personal data went in, assess it as a possible personal data breach. The Dutch data protection authority has stated that when employees enter personal data into a chatbot on their own initiative and against the employer’s agreements, that is a data breach, and that notifying the regulator and the affected people is in many cases mandatory (Autoriteit Persoonsgegevens, 6 August 2024). Under UK GDPR, notifiable breaches go to the ICO within 72 hours of becoming aware, where feasible, and every breach must be recorded even when it is not reported (ICO). For US covered entities, an impermissible disclosure of protected health information is presumed to be a breach unless a four-factor risk assessment shows a low probability of compromise (HHS). Bring in legal and privacy on day one for tier 1 tools, because the 72-hour clock can start while you are still sorting the list. The full incident steps, by data type, are in what to do when an employee pastes sensitive data into ChatGPT.
Why blocking AI domains alone fails, and what to do instead
A domain blocklist answers one question, whether a device on your network can reach a hostname. Shadow AI escapes it in at least five ways.
- Personal and corporate accounts share a domain. Google’s admin documentation says a URL-filtering proxy “won’t work” for blocking consumer accounts because legitimate traffic from a managed account “goes to the same URL” (Google Workspace Admin Help). ChatGPT has the same shape.
- Laptops leave the network and phones were never on it. A block on the office proxy does not follow a laptop to home Wi-Fi unless traffic is forced through a client, and it never sees a personal phone.
- Desktop apps and command-line tools travel differently. OpenAI’s documentation for its workspace header notes that the ChatGPT desktop and mobile apps use certificate pinning, so the header needs pinning exceptions pushed by device management, and that logged-out ChatGPT needs a separate block (OpenAI Help Center, October 2026). Coding agents call APIs from a terminal.
- New tools appear faster than lists update. Cornell’s IT team warns that some AI notetaker vendors now use “multiple or alternate domains to circumvent attempts to block by domain” (IT@Cornell).
- Bans push the work underground. In the University of Melbourne and KPMG global study of 48,340 people, uploading sensitive company information to public AI tools was reported by 67% of employees whose organization had banned generative AI, against 33% where there was no policy; the authors wrote that “outright bans may be ineffective” (University of Melbourne and KPMG, 2025).
Most security teams do not rely on strict blocking anyway. In the CSA survey, only 22% strictly block data going to GenAI, even though GenAI was the channel most often ranked as the top risk.
Single-browser controls have a similar edge. Microsoft’s Purview documentation for browser data security, updated in October 2026, applies its policies for a listed set of consumer AI apps inside Edge for Business, version 144 or later, on Intune-managed Windows 10 and 11 devices, and users who open those apps in another browser are blocked and sent to Edge. Work done in a desktop app or a terminal sits outside a browser control, and any control keyed to sensitivity labels acts only on data someone has labeled.
To block unapproved AI tools while keeping the sanctioned ones, layer the controls. Use vendor tenant controls so only your workspace loads on managed networks: for ChatGPT Enterprise workspaces, OpenAI’s ChatGPT-Allowed-Workspace-Id header filters out “any other workspaces (including personal workspaces)”, and Google’s X-GoogApps-Allowed-Domains header does the same for Google sign-ins, both through a TLS-inspecting proxy. Block the tier 1 tools that have no business tier. Then put an account-aware control on the laptop, so a paste into a personal account is caught wherever the laptop is. How to block ChatGPT at work has the configuration steps for firewalls, browsers, desktop apps and APIs.
The long tail that is not a chatbot
Many entries on a discovery list are not chat websites at all, and several never show up on a domain report.
AI features inside SaaS you already pay for. Meeting summaries, document assistants and CRM agents switch on inside approved apps, and the processing happens on the vendor’s servers. The control is the admin console and the contract. How to control AI features in SaaS apps lists the settings by vendor.
Browser extensions. In December 2025, Ars Technica reported that eight extensions with more than 8 million installs were capturing complete AI conversations from ChatGPT, Claude, Gemini and five other platforms and sending them to the extension maker (Ars Technica, 17 December 2025). A VPN or ad blocker can read every prompt without being an AI tool on anyone’s list. Chrome’s admin settings support “Block all apps, admin manages allowlist” for managed browsers (Google Chrome Enterprise Help).
AI notetakers. A bot is usually authorized by one attendee, joins meetings hosted by others, and records whatever is said. Sanction one notetaker with admin controls and keep the rest out of company meetings. Are AI notetakers a security risk? covers the meeting-platform settings.
Coding assistants and agents. These read repositories, environment files and terminals, and they reach the model through APIs. Sanction the enterprise tier and restrict login to your organization. See are AI coding assistants safe for source code? and how to secure AI agents on employee computers.
OAuth-connected apps. A user who clicks “Sign in with Google” can grant an AI app standing access to mail and files. In Google Workspace the default setting for unconfigured apps is “Allow users to access any third-party apps”, and admins can restrict that to reviewed apps (Google Workspace Admin Help). OAuth reviews only catch tools tied to a corporate identity; an app opened with a personal Gmail account never appears there.
Personal and corporate accounts of the same tool
The hardest line on the list is the one that says “ChatGPT” or “Claude” when you already license it. Part of that traffic is your company workspace and part is personal accounts, and a domain report cannot separate them. Tenant headers separate them only on traffic that crosses your proxy. The fix combines moving people into the company account, giving them the features that drove them to personal accounts, and an account-aware control on the laptop for the data that must not go to a personal account. Why employees keep using personal AI accounts after you approve a tool goes through the reasons and the fixes.
Tell staff before you block: a no-blame message and a self-report window
Most people on the list were trying to get work done. Hamidi put it plainly: “What I don’t want to do is punish people for using AI for increasing their productivity” (CSO Online, March 2026). TechTarget describes an AI amnesty as a voluntary, time-limited program in which employees disclose unauthorized AI use without punishment, run by a committee drawn from legal, HR, IT and risk management (TechTarget, 15 April 2026).
A message that works says five things:
- We found AI tools in use that we have not reviewed, and we know most of you used them to work faster.
- Here are the approved tools, and how to get access today.
- Until a stated date, tell us which AI tools you use, what you put into them and why, through this form. Nobody will be disciplined for what they disclose in this window.
- If you put customer, patient, employee or financial data, source code or credentials into an unapproved tool, tell us now so we can limit the damage.
- After the window, these tools will be restricted, and here is how to ask for a new one.
Ask what data went in, and treat the answers as input to the tier 1 list. Reserve formal consequences for repeat behavior after the window and for deliberate removal of data, and agree that line with HR before the message goes out.
Who decides: security, IT, legal and privacy, and the business owner
No single team should own the call. In the CSO Online piece, Hamidi describes a review covering information security, legal and data privacy, plus a return-on-investment check, before a tool is approved or rejected. IBM’s 2026 report found only 19% of organizations coordinated their AI governance and security efforts.
| Role | Owns |
|---|---|
| Security | The inventory, the data tiers, blocks and endpoint controls, incident triage |
| IT | Licenses, SSO, tenant headers, browser and extension policy, SaaS admin settings |
| Legal and privacy | Vendor terms, data processing agreements, breach assessment and notification |
| Business owner of each tool | Whether the need is real, who uses it, funding for seats |
| HR | The staff message, the self-report window and consequences after it |
Give the group a standing meeting and a turnaround target for requests, so decisions do not queue behind the next incident.
How to avoid shadow AI in enterprises after the first cleanup: shadow AI management best practices
A one-time cleanup goes stale, because new tools launch, existing ones rename themselves, and approved apps switch on new AI features. Keeping the list current takes four habits.
- A request path with a published turnaround. NIST’s Generative AI Profile recommends maintaining “approved GAI technology and service provider lists” (GV-6.1-007) and defining inventory exemptions for GenAI “embedded into application software” (GV-1.6-002) (NIST AI 600-1, July 2024). Log every request in the same inventory as the discovery results.
- A re-check cadence. Re-run log, OAuth and expense queries monthly and compare each run with the last. Re-read the data terms of every sanctioned tool at renewal and whenever the vendor announces a change; Anthropic’s August 2025 consumer change, which introduced five-year retention for users who allow training, is the kind of update that changes a tier.
- An AI acceptable use policy that names tools and data. List the approved tools, the data each may receive and the account people must use. How to write an AI acceptable use policy includes a template.
- A few numbers reported each cycle. Track the share of AI use on company accounts, the number of tools receiving tier 1 data and the time to approve a request.
How Jazz turns the tool list into evidence
Jazz collects its evidence on the laptop itself, through a single endpoint agent on managed Windows and macOS devices, so a portable browser or the ChatGPT desktop app does not create a gap.
That changes what the triage list contains. In one environment where security had sanctioned a handful of GenAI tools, Jazz’s agent found 1,223 in use (Security That Says Yes, September 2026); another customer discovered 447 GenAI tools with no prior visibility into what was flowing into them. For each one, the question that matters is what data went in, from which account.
Melody, Jazz’s agentic investigator, answers it. Each data movement is examined across four dimensions: the data itself, the systems it moved between, the person and their role, and whether it fits a known business process. A sanctioned corporate GenAI account and a personal account on the same service are different destinations in that analysis. Tomer Maman, CISO of Similarweb, says on Jazz’s homepage that Jazz lets his team “distinguish sanctioned enterprise use from personal exposure.” Security receives a pre-investigated incident that says what was shared, by whom, into which tool and account, which is the input tier 1 triage and the breach assessment both need.
When Jazz does intervene, the response matches the risk: a prompt pointing the user to the approved tool, a request for justification, escalation for review, or a targeted block. That is the “move users” decision enforced at the moment of the paste, without blocking the domain for everyone.
Jazz’s coverage has limits. It covers managed Windows and macOS laptops with the agent installed. It does not see personal phones, unmanaged or personal computers, or AI processing that happens server-side inside a SaaS platform, such as an assistant summarizing channels inside the vendor’s cloud. Those need the SaaS admin settings and identity controls: OAuth consent restrictions, tenant restrictions and the AI settings in each app’s admin console.
Frequently asked questions
We just discovered employees using 40+ unapproved AI tools. Where do we even start?
Rank the tools by what company data goes into each one and how many people use it. Act first on the few receiving regulated or confidential data, such as customer records, source code or credentials. For each tool, decide whether to sanction it, move its users to your approved tool, replace it or block it.
Why is shadow AI important to address?
Because it now shows up in breach data. IBM's 2026 Cost of a Data Breach Report found shadow AI involved in 43% of security incidents, up from 20% a year earlier. Those breaches averaged USD 5.39 million, and data loss or compromise followed in 49% of them.
We tried blocking ChatGPT and employees found workarounds. What should we do instead of blocking AI tools?
Give people an approved tool and move them onto it, then restrict personal accounts for sensitive data. In the University of Melbourne and KPMG study, 67% of employees whose organization banned generative AI reported uploading sensitive company information to public AI tools, against 33% with no policy. Keep hard blocks for tools with no business tier that receive sensitive data.
How do I block unapproved AI tools but still allow the ones we've sanctioned?
Layer the controls. Vendor tenant headers, such as OpenAI's ChatGPT-Allowed-Workspace-Id for ChatGPT Enterprise, let only your workspace load through a TLS-inspecting proxy. Block tools with no business tier that receive regulated data, and add an account-aware control on the laptop for off-network use. How to block ChatGPT at work has the configuration steps.
What should we do about company data employees already put into unapproved AI tools?
Find out what went in and from which account. Then read the vendor's terms for that plan, since consumer plans may train on chats and keep reviewed or opted-in data for years. Ask the account holder to delete it, record that, and have legal and privacy assess whether it is a reportable breach.
Can we make employees delete company data from their personal ChatGPT or Claude accounts?
Not directly. For a personal account, the account holder controls deletion. Claude Enterprise admins can claim and migrate personal accounts on a verified company domain, but ChatGPT Business admins cannot force a personal workspace to merge or delete. Accounts opened with a personal email stay outside every employer feature.
Should we offer an amnesty to employees who used unapproved AI tools?
Yes. Announce a time-limited window in which staff can disclose which AI tools they used and what data went in, with no discipline for what they disclose. Publish the approved tools at the same time. The disclosures feed your tier 1 list, and formal consequences are reserved for repeat behavior after the window and deliberate data removal.
Who should decide which AI tools get approved?
A small group, not security alone. Security owns the inventory and controls, IT owns licences and admin settings, legal and privacy own vendor terms and breach decisions, and the business owner of each tool confirms the need and funds seats. HR runs the staff message and any consequences.
How can we tell employees' personal ChatGPT accounts from our company workspace?
A domain report cannot, because both use the same domain, and tenant headers work only on traffic that crosses your proxy. On managed Windows and macOS laptops, Jazz's agent sees the paste or upload at the operating-system level and treats the company workspace and a personal login as different destinations, on any network. It can then point the user to the approved workspace, ask for a justification or apply a targeted block.
Is there a way to see what data employees put into unapproved AI tools, not just which sites they visited?
Only an endpoint control or the employee's own report shows what was pasted or uploaded. Jazz captures GenAI prompts, copy and paste and file uploads on managed Windows and macOS laptops in browsers, desktop apps and command-line tools, and Melody returns a pre-investigated incident saying what was shared, by whom, into which tool and account. Phones and unmanaged computers are outside it.
Does Jazz cover phones, personal laptops and AI built into SaaS apps?
No. Jazz covers managed Windows and macOS laptops with its agent installed, where it sees GenAI prompts and uploads in any browser and in desktop AI apps, without an extension. Personal phones, unmanaged computers and AI processing inside a SaaS vendor's cloud need SaaS admin settings, OAuth consent restrictions and tenant controls.
How do we keep shadow AI from coming back after the first cleanup?
Publish an approved tool list with a request form and a turnaround target. Re-run log, OAuth and expense queries monthly, and re-read each sanctioned vendor's data terms when they change. Keep an AI acceptable use policy that names the approved tools, the data each may receive and the account to use.



