The useful unit of AI implementation is not the tool, department or job. It is a decision that happens often enough to examine and clearly enough to improve.
MIT’s 2025 study “The GenAI Divide: State of AI in Business” reported that 95 percent of enterprise generative AI pilots produced no measurable P&L impact, across 52 executive interviews, 153 survey responses and 300 public deployments.
Treat that headline with the caution it deserves. It is a small sample and it has been reasonably criticised for how widely it has been repeated. The finding underneath it is the durable part, and it matches what we see: the failures were not model failures. They were organisations unable to attach the tool to a real process.
Which is a design problem, and design problems have a unit. This one is the decision.
Most company AI strategies start several levels too high. They announce that the business will become AI-first. They appoint a working group. They list approved platforms. Then they ask every department to find a use case.
That order produces activity, not direction.
A company does not experience work as a collection of technologies. It experiences work as a sequence of decisions:
- Is this lead qualified?
- Which campaign should receive the next euro?
- Does this invoice match the agreed scope?
- Is this asset ready to publish?
- Which customer message needs a human response?
- Has this product record changed enough to require a new version?
Some decisions are frequent and reversible. Some are rare and consequential. Some rely on clean evidence. Others depend on context that has never been written down.
AI belongs in different roles across those decisions. Until the decisions are mapped, the company cannot know which role is appropriate.
Why a list of use cases is not enough
“Use AI for customer service” is not a use case. It is a department with dozens of different decisions hidden inside it.
A support workflow may include identifying the customer, classifying the request, retrieving the relevant policy, drafting an answer, deciding whether an exception is justified, issuing a refund, detecting possible abuse, and recording what happened.
Those steps do not carry the same risk. Drafting a response from an approved policy is different from authorising a refund. Tagging a ticket is different from deciding that a customer is acting fraudulently.
If the whole workflow is labelled “customer service automation,” the implementation has already hidden the decisions that need separate controls.
The ILO’s 2025 global index estimated that 25 percent of global employment falls within occupations potentially exposed to generative AI. It found transformation more likely than full replacement, because complete job automation remains limited and many tasks still require human involvement.
Jobs are too large a unit. Departments are too large a unit. Decisions are small enough to design.
The Recurring Decision Map
Before selecting a platform, create a Recurring Decision Map. It is a register of decisions that happen repeatedly across the company. Each decision receives one row and nine fields.
1. Decision
Write the decision as a question with a clear outcome. Weak: “manage incoming leads.” Useful: “does this enquiry meet our minimum criteria for a sales conversation?”
The question format exposes what is being decided. It also makes disagreement visible.
2. Trigger
What causes the decision to happen? A form submission, a customer reply, a stock threshold, a campaign reaching a spend limit, a document entering a folder, a scheduled review date.
If there is no reliable trigger, the process may not be ready for automation.
3. Owner
Who is accountable for the result? This must be a role or a named person, not “the business.” AI can perform part of the work. It cannot absorb accountability.
4. Evidence
What information should the decision use? List the actual sources: contract, product record, customer history, campaign report, approved policy, inventory system, previous decision log.
Then mark each as structured and current, available but inconsistent, mostly held in messages, dependent on personal memory, or missing.
AI does not repair weak evidence by reading it faster.
5. Frequency
How often does the decision occur? Use a real count: times per day, per week, per project, per 1,000 transactions.
Frequency turns repeated effort into measurable operating cost and provides more examples for testing. It does not by itself create a sufficient evaluation sample, especially when consequential failures are rare.
6. Consequence
What happens if the decision is wrong?
- Low: inconvenience or minor rework
- Moderate: customer friction, wasted spend or missed time
- High: material financial, contractual or reputational harm
- Critical: safety, rights, employment, health or irreversible loss
Do not confuse a common decision with a low-risk decision.
7. Reversibility
Can the action be undone? Drafting an internal summary is highly reversible. Sending it to a client is less reversible. Publishing it, transferring money or deleting a record may be effectively irreversible.
Reversibility often matters more than confidence. A system can safely receive more autonomy when a mistake is cheap to detect and undo.
8. Exception rate
How often does the normal rule fail? A process that looks repetitive may rely on human judgment every third time. Record the percentage of cases requiring an exception, the most common exception types, who currently resolves them, and whether the resolution becomes a reusable rule.
Do not estimate from memory if a sample can be reviewed.
9. Current decision time
Measure total elapsed effort: gathering information, waiting for missing information, making the decision, checking it, correcting it, recording it. The visible click is rarely the whole cost.
Sort decisions into four roles
Human-only. Use when the decision is highly consequential, difficult to reverse, dependent on unrecorded context or inseparable from human responsibility. The system may still organise evidence, but it does not recommend or act.
AI-assisted. AI retrieves, compares, drafts or flags. A person reviews the evidence and makes the decision. This is the correct starting role for many knowledge-work decisions.
Rule-automated. Software executes a deterministic rule. The rule, inputs and exception path are explicit. Not every useful automation needs a language model.
Bounded delegation. An AI system may choose and act inside a defined permission envelope, with limited tools, an explicit stop condition and a named escalation path. This role should be earned through evidence. It should not be the default because the platform supports agents.
The first-pass filter
Five questions find the first implementation candidates:
- Does the decision happen at least weekly?
- Are the required inputs available before the decision?
- Can a correct result be described or evaluated?
- Is a mistake detectable?
- Is the first version reversible?
In our first-pass screening heuristic, five yes answers suggest a strong candidate for deeper evaluation. Three or four suggest assistance or process cleanup. Fewer than three suggest the process may still be too ambiguous. These cutoffs are our operating rule, not validated predictive thresholds.
This is a filter, not a compliance assessment. Decisions affecting employment, access to essential services, health, credit, safety or fundamental rights require specific legal and domain review regardless of their score.
A 90-minute mapping session
The first map does not require a transformation programme.
Minutes 0 to 15, collect verbs. Ask each participant to list the repeated decisions they make: approve, reject, classify, choose, prioritise, compare, escalate, publish, refund, order.
Minutes 15 to 35, rewrite them as questions. Turn every vague activity into a decision with an observable outcome.
Minutes 35 to 55, add evidence and ownership. Identify the source used and the person accountable.
Minutes 55 to 75, score consequence, reversibility and exceptions. Disagreement is useful. It reveals where the company has no shared operating rule.
Minutes 75 to 90, select three candidates. One low-risk assisted decision, one deterministic automation, and one decision that should remain human. The third matters. A credible AI programme needs explicit exclusions.
What the map reveals before any AI is added
The exercise exposes problems that are not AI problems: two teams using different definitions for the same outcome, nobody owning the final decision, evidence stored inconsistently, exception handling that is invisible, real work happening in private messages, people checking each other’s work without recording why, a supposedly automatic process depending on one experienced person.
These findings are not a delay. They are the implementation.
NIST’s AI Risk Management Framework starts from a similar premise: the business context, the specific tasks, the intended scope, human oversight and third-party dependencies should be mapped and documented. The framework does not begin with model selection.
The EU AI Act adds a practical reason for European companies to take this seriously. Its AI-literacy obligation has applied since 2 February 2025. Under Regulation (EU) 2026/1744, providers and deployers must take measures supporting AI literacy among staff and others operating AI systems on their behalf, though they do not have to guarantee a specific level for each individual. The Act’s general application date remains 2 August 2026.
Training people to use a chatbot is not enough if the company has not defined what they are using it to decide.
How we run this
We do not begin an AI engagement by asking a client which model it wants. The first deliverable is a Recurring Decision Map: what the business decides, how often, using which evidence, with what consequence, under whose authority, and with which exceptions.
Only then does Origin, Scale or Merch become the relevant operating lens. The division follows the decision. The technology follows the division. The accountability stays with us and with the people operating the system.
An AI strategy can be written without understanding the work. A map of recurring decisions cannot.
If you want to run the 90-minute session on your own operation, tell us what your team decides most often and we will tell you which three to start with.
Sources
MIT, “The GenAI Divide: State of AI in Business 2025”, reported coverage, and a critical reading of the same study.
NIST AI Risk Management Framework Core.
ILO, Generative AI and Jobs: A Refined Global Index of Occupational Exposure.
OECD, AI adoption by small and medium-sized enterprises.
European Commission, AI Act implementation timeline.
Regulation (EU) 2026/1744 amending the AI Act.
Current to 30 July 2026. This article provides an operating framework, not legal advice.
