Fire In Belly

Shadow AI: Turning Unsanctioned Use Into a Sanctioned Internal Workflow Backlog

14 min read

Shadows of palm fronds cast across a grey textured wall

Your security team runs a two-week discovery sprint and comes back with a list: 40 AI domains, a dozen browser extensions, three personal subscriptions on expense claims. Leadership buys an enterprise chat licence, blocks the rest, and announces a policy. Six months later the licence is half used, and the finance team is still pasting supplier data into a personal chatbot every Monday morning.

That is the common way shadow AI programs stall. Shadow AI is the use of AI at work that sits outside your approved systems and processes. The tool list tells you where the risk is, but it does not tell you what to build. This guide shows how to turn discovered shadow AI into a ranked backlog of sanctioned replacements: inventory by task, sort each task into one of three outcomes, rank the ones that need a real workflow, and publish enforcement only after the replacements exist.

What shadow AI guidance already covers

The definitions and the risk case are settled. The UK NCSC defines shadow AI as AI technology not captured in an organisation's approved systems and processes, cites one study in which 71% of employees reported using AI tools their employer had not approved, and says the use is unlikely to disappear. It tells organisations to understand why staff turn to these tools so they can provide secure alternatives.

Detection is well documented too. Flexera's guide walks through network, SaaS posture, identity and OAuth, endpoint, DLP, gateway, and agent monitoring, and is candid about their limits: a network log can show a connection to an AI provider without showing which app made the call or what data went with it. CyberReplay sets out a 90-day plan that starts with a 14-day discovery sprint over SSO, DNS, and cloud audit logs.

The per-tool decision is covered as well. Easy Cyber Protection recommends an anonymous three-question survey (which tools, for what tasks, with what kind of data), then a call per tool: approve, replace, or block. Barracuda frames the same choice as approve, deny, or redirect. Forcepoint publishes an acceptable-use template that lists approved tools by tenant and account type.

Those pages cover detection and policy well. They leave open what replaces the use once you have found it.

Why a tool list cannot tell you what to build

Every discovery method above records a tool: a domain, an app, an account, an OAuth grant. The replacement decision depends on something else, which is the task. Two people hitting the same chatbot domain can be doing completely different work. One is rewording an email. The other exports 200 rows from the finance system each week, asks the chatbot to classify them, and pastes the result back into a spreadsheet that feeds payment runs.

An enterprise chat licence replaces the first use well. It does nothing useful for the second. The copy-paste is still manual, the export still leaves the system of record, and the person still has to check the output by hand. The tool is now approved, but the work is unchanged.

Inventory shadow AI by task and by where the output goes next, not by tool: a DNS or CASB log shows which AI service was reached, but only the task record shows whether an enterprise chat licence will replace the use or whether a copy-paste round trip between a system of record and a chatbot needs a built workflow. Flexera is right that policy should be written around use case and data flow, not only the product name. Apply the same idea to the inventory: record what goes in, what comes out, and where the output goes next. The output destination is the field that separates a licence decision from a build decision.

The NCSC's Shadow IT guidance points the same way. It lists AI that helps with rewriting documents, compiling information, and summarising meetings among the unmet needs that push staff to unofficial tools, and it tells security teams to bring shadow IT above board by addressing the underlying user need, including by migrating data into supported platforms. A need is a task, not a product.

There is evidence that tool-for-tool swaps disappoint. The MIT NANDA State of AI in Business 2025 report, which is based on interviews and surveys with stated sample limitations, found workers at over 90% of surveyed companies using personal AI tools for work while only 40% of companies had bought an official LLM subscription. It also found users rejecting enterprise tools that felt static next to consumer chat. One practitioner on Hacker News describes a university's enterprise ChatGPT rollout with low adoption and retention, which the author attributes largely to features it lacked. That is one author's report, but it is the failure a tool-level replacement invites.

Run a no-blame task inventory before you announce enforcement

The guidance pulls in two directions on sequencing. The NCSC's shadow IT guidance says to take a positive, no-blame approach, because if you blame or punish staff, their peers will be reluctant to tell you about their own unsanctioned practices. Easy Cyber Protection says to promise no punishment for reporting. Meanwhile, Forcepoint's template says violations may result in disciplinary action up to and including termination, and CyberReplay recommends setting disciplinary consequences in the acceptable-use policy.

Both can be right, but not at the same moment. Our recommendation, reasoning from those sources: run the no-blame task inventory before you publish disciplinary consequences, because once a policy says violations can end in termination, the people whose tasks you most need to understand stop telling you about them. Run it under an explicit amnesty and publish enforcement afterwards. If the termination clause goes out in week one, the finance analyst with the Monday export has every reason to keep quiet, and that export is exactly the task you most need on the list. The 2024 Work Trend Index found that 78% of AI users bring their own AI tools to work and that 52% are reluctant to admit using AI for their most important tasks. People already hide their AI use, and a published threat gives them another reason to.

Keep the amnesty narrow and honest. It covers past use disclosed during the inventory window. It does not cover future use of restricted data in unapproved tools, and it does not stop you rotating a credential someone admits to pasting. Say both things up front.

The inventory fields

Extend the three-question survey with fields that answer the build question. Ask per task, not per person.

FieldExample answerWhy it matters
TaskSummarise supplier invoice exceptionsThe unit you will replace
Input sourceExport from the finance systemShows whether data leaves a system of record
Data classSupplier names, amounts, bank detailsSets the floor for any replacement
Output destinationPasted into the payments spreadsheetSeparates a licence decision from a build
Frequency and timeWeekly, about 90 minutesFeeds the ranking
People doing itTwo analystsFeeds the ranking and names pilot users
Tool usedPersonal chatbot accountNeeded for risk, least useful for the build

Combine the answers with the technical signals your IT team already has. The survey finds tasks a log cannot describe, and the log finds tools nobody mentioned.

Sort every shadow AI task into one of three outcomes

Each task lands in one of three places. Write down the reason for every call, because you will revisit them.

A licence is enough

The task is one person, one prompt, and the output goes into that person's own writing or thinking: drafting an email, rewording a paragraph, brainstorming, summarising a public document. A sanctioned chat tool with the right account type and data terms replaces it. The MIT NANDA report found that most users already prefer AI for this kind of quick work, so expect quick uptake. Your job is to make sure the approved tool is at least as capable as the one it replaces, or people will keep the old tab open.

A built workflow is needed

The task has any of these signs:

  • Data is copied out of a system of record and the result is pasted back into a system.
  • It repeats on a schedule or per item: every invoice, every ticket, every Monday.
  • The output needs a fixed structure that something downstream depends on.
  • More than one person does the same task in slightly different ways.

The first sign is the strongest. A copy-paste round trip means the system already holds the input and expects the output, and only a human with a chatbot is connecting them. A licence leaves that round trip in place. A built workflow reads the source directly, applies the model step under your controls, and writes the result where it belongs, with logging and review. In this category, shadow AI shows you automation demand that people have already proven by doing the work, and that your official roadmap may have missed.

It stays manual or blocked

Some tasks should not go into any AI tool yet: restricted data with no acceptable processing terms, decisions with legal weight and no review path, or tasks too rare to justify anything. Block what needs blocking and explain why. As Easy Cyber Protection puts it, an explained block is followed, and a silent one is circumvented.

Rank the workflow candidates

You will usually have more built-workflow candidates than build capacity. A simple, visible scoring rule beats a debate. This one is our recommendation, not an industry standard; adjust the weights to your risk appetite.

hours_per_month = frequency_per_month * minutes_per_run * people / 60
exposure_weight = 3 if data_class in {restricted, regulated}
                  2 if data_class == confidential
                  1 otherwise
round_trip      = 2 if output is pasted back into a system of record else 1
priority        = hours_per_month * exposure_weight * round_trip

Exposure counts twice in effect: once because the shadow use is risky today, and again because a built replacement removes the export entirely. A low-hour task with restricted data can outrank a high-hour task on public data, which is usually the right call.

Then apply two gates before anything enters the backlog. First, the task owner agrees the description is accurate. Second, someone can name the system the workflow would read from and write to. If nobody can, the task is not ready to build; it needs process mapping first.

A worked example: the supplier-invoice exception summary

Here is an illustrative case built from the pattern above; the numbers are hypothetical.

Two accounts payable analysts each spend about 90 minutes every Monday on supplier invoice exceptions. They export last week's unmatched invoices from the finance system, paste them into a personal chatbot, ask it to group the exceptions by likely cause and draft a note per supplier, then paste the groups into the payments spreadsheet and send the notes by email.

The inventory captures it as weekly, two people, confidential and some restricted data (bank details appear in a few rows), output pasted back into a spreadsheet that feeds payment runs. Under the ranking rule that is about 13 hours a month, an exposure weight of 3, and a round-trip multiplier of 2, so a priority of roughly 78. A colleague who rewords supplier emails for 20 minutes a day scores far lower and gets the licence.

The built replacement reads unmatched invoices from the finance system on a schedule, strips fields the model does not need, groups exceptions with a model step that returns a fixed structure, and writes the groups into a review queue. The analysts approve or correct each group, and approved notes go out from the shared mailbox. Bank details never leave the finance system. The Monday round trip disappears, and the analysts spend their time on corrections instead of copying.

Recruit the people already doing the work

The analysts in that example already worked out that the task could be done with AI, and they know where it goes wrong. The MIT NANDA report observed that many of the strongest deployments began with power users who had already experimented with consumer tools and became early champions of sanctioned solutions. It also noted organisations learning from shadow usage before procuring enterprise alternatives.

The employees already doing a task in a personal AI tool are the best pilot users and the best source of test cases, but their chat history can hold the data you are trying to protect, so collect the shape of the task and synthetic examples rather than raw transcripts. Flexera notes that a compromised personal AI account can expose prior conversations containing business context people forgot they typed. Put those users in the pilot and ask them for their edge cases. What to collect: the steps they take, the instructions that worked, the mistakes they learned to catch, and examples rebuilt with synthetic values. Then ask them to clear the old history from the personal account once the sanctioned version is live.

That gives you a first draft of the workflow spec and a starter set of test cases, without importing a second copy of the data exposure into your own repository.

Then publish the enforcement you need

Once the inventory is done and the first replacements are live, enforcement has something behind it. Publish the approved list by tenant and account type, as Forcepoint's template suggests, and the restricted-data rule. Add the blocks for the third bucket and the technical controls your security team chose from the detection guides above.

Keep a request path open and fast. The NCSC's shadow IT guidance says a slow or ineffective request process encourages people to implement their own solutions. If requests sit in a queue for weeks, expect the next round of shadow AI to start there.

Failure modes and tradeoffs

The licence is weaker than the personal tool. People keep using the personal one. Compare capabilities before you buy, and treat a gap as a defect to fix, not a training problem.

The inventory turns into a confession booth. If the form asks for names next to restricted-data admissions, response rates drop. Collect tasks anonymously by team, and follow up only with volunteers.

The backlog becomes a wish list. Every team adds a task, and nothing ships. Hold to the two gates and build the top one or two first.

Blocking before replacing. You lose visibility and the work moves to phones. Easy Cyber Protection and Flexera both warn about this.

Building where a licence would do. A custom workflow for a one-person writing task costs more than it saves. The three-way sort exists to stop this.

How to verify the backlog is working

Check four things monthly for each replaced task:

  1. The task owner confirms the old personal-tool route is no longer used for that task.
  2. Rows exported from the source system for that task fall to zero.
  3. Review corrections in the new workflow are logged and trending down, or the reason is understood.
  4. The shared monthly hours for the task are lower than the inventory baseline.

For adoption measures after launch, see our guide on measuring the ROI of AI workflow automation. For the security controls around any internal assistant you approve, see AI chatbot security for internal deployments and AI data privacy for internal AI tools. For review design in the workflows you build, see human-in-the-loop AI governance.

Next action

Pick one team this week and run the shadow AI task inventory with them under an explicit amnesty. Use the seven fields above, sort each task into licence, built workflow, or manual, and score the workflow candidates. If the top candidate has a copy-paste round trip into a system of record, that is your first build.

If you want help mapping those tasks and ranking them against your systems, our AI audit does exactly that in a week, and AI workflow automation builds the sanctioned replacement.


About Fire In Belly: Independent senior engineering from Tallinn, Estonia. We design and build AI workflow automation, internal tools, and custom software with the task-level shadow AI inventories, scored backlogs, and sanctioned replacement workflows described above, at published fixed prices. Schedule a call to discuss your next project.