Short answer: Pick a workflow with one owner, enough volume, a baseline you can measure and a clear point where a person approves the result. Score eight to fifteen candidates with the people who do the work, measure today’s time and errors, then build the winner into the tools the team already uses.
The first workflow matters more than its size suggests. It sets who owns AI work and how results get measured.
In McKinsey’s 2026 State of AI survey, 37 percent of respondents attribute at least some EBIT impact to AI. Among the roughly 6 percent McKinsey calls high performers, nearly three-quarters report fundamentally redesigning workflows because of their AI use, against one-quarter of everyone else. The survey is self-reported, and correlation is not cause. The lesson still holds: redesign one real workflow rather than add a tool and hope.
Here is the method in six steps. You can start it this week with a spreadsheet and the people who do the work.
1. List candidates with the people who do the work
Hold a one-hour session with the people who run the process every day. Do not ask where AI could go. Ask where the time goes.
What do you read, copy, check or rewrite every day?
Where do cases sit waiting for someone?
What gets sent back, corrected or chased?
What would you stop doing first if you had a free afternoon each week?
Write each answer as a workflow with a start and an end. “A supplier invoice arrives and is matched against the order and the goods receipt” is a candidate. “AI for finance” is not. Aim for eight to fifteen candidates. Typical ones look like this:
an incoming supplier invoice matched against an order;
a customer question answered from approved product information;
a retailer’s new-item form, filled in from your own product data;
a batch record or deviation report prepared for quality review.
2. Score each candidate
Score every candidate from 1 to 3 on the eight criteria below. A 2 sits between the two descriptions. Three criteria are gates: a 1 on any gate means the candidate waits, whatever its total.
Volume and frequency. 1: a few times a month. 3: many times a day.
Time per case. 1: a few minutes. 3: half an hour or more.
Text-heavy inputs. 1: structured data already in a system. 3: documents, emails, forms, free text.
Clear definition of “done”. 1: two experts often disagree. 3: a checklist or rule decides.
Cost of error and reversibility (gate). 1: costly and hard to undo. 3: cheap, or caught before it leaves.
Data access (gate). 1: no access yet. 3: access through a system you control.
Natural approval point. 1: none exists. 3: an existing sign-off step.
One owner (gate). 1: nobody, or a committee. 3: a named person with time and authority.
The example below is hypothetical and only shows how the scoring works. A is invoice matching, B a deviation report draft, C retailer item-setup forms.
Criterion | Scores A · B · C |
|---|---|
Volume and frequency | 3 · 2 · 2 |
Time per case | 2 · 3 · 2 |
Text-heavy inputs | 3 · 3 · 3 |
Clear “done” | 3 · 2 · 2 |
Error cost (gate) | 2 · 2 · 2 |
Data access (gate) | 2 · 2 · 1 |
Approval point | 3 · 3 · 2 |
One owner (gate) | 3 · 3 · 2 |
Total (max 24) | 21 · 20 · 16 (C waits) |
Favor the highest total that passes all three gates. Break ties on the clearest approval point. In the example, invoice matching and the deviation report both pass. Retailer item setup fails the data-access gate and waits.
Text-heavy inputs earn a row because that is where a language model adds something rules cannot. OpenAI’s practical guide to building agents points to complex judgment, hard-to-maintain rules and “heavy reliance on unstructured data.” Where a use case does not clearly fit, it says, “a deterministic solution may suffice.” A candidate that scores 1 here may be a job for ordinary automation.
3. Measure the baseline before you build
You cannot show a change you did not measure. Before anyone builds anything, record four numbers for the chosen workflow over two to four weeks of normal work:
Cycle time: from the moment a case arrives to the moment it is done.
Hands-on time: minutes of human work per case.
Rework or error rate: the share of cases corrected, sent back or escalated.
Backlog: cases waiting, and the age of the oldest.
Also set aside real past cases with known correct outcomes as a test set. OpenAI’s guide advises teams to “set up evals to establish a performance baseline” for the system itself. Your past cases are the raw material.
McKinsey’s 2026 survey found high performers twice as likely as others to report defined processes to measure the impact of their AI initiatives.
4. Draw the boundaries
Write one page with four headings. For invoice matching, it might read:
What the system does: reads the invoice, extracts the fields, matches them against the order and goods receipt, flags differences, and drafts a note to the supplier.
What a person decides: approves payment, sends the supplier note, and changes supplier master data.
When it escalates: a mismatch above the agreed tolerance, a new supplier, a missing receipt, or a document it cannot read.
Who can stop it: the named owner, and how.
For a customer-question workflow, the same page might say that the system drafts answers only from approved product information. A person approves anything about safety, health or refunds. Questions outside the approved content go straight to the team.
OpenAI’s guide names two triggers for handing control to a person: exceeding failure thresholds, such as repeated retries, and high-risk actions. “Actions that are sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent’s reliability grows.” McKinsey’s 2025 survey found high performers more likely to have “defined processes to determine how and when model outputs need human validation to ensure accuracy.” Decide this at design time, not after the first mistake.
5. Build it into the tools people already use
The workflow should live where the work already happens: the inbox, the ERP, the document system, the ticket queue. A separate screen that people must remember to open is an easy way to lose them.
MIT NANDA’s GenAI Divide report found that most enterprise-grade AI tools the organizations evaluated never reached production. Most, it says, fail “due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.” Read it with care: these are preliminary findings from 52 interviews and 153 survey responses, which the authors call directionally accurate. The direction is consistent with McKinsey’s finding on workflow redesign.
Then start with the simplest design that works. Anthropic’s Building effective agents separates workflows, where models and tools follow “predefined code paths,” from agents, which “dynamically direct their own processes and tool usage.” It recommends “finding the simplest solution possible, and only increasing complexity when needed,” and notes that agents’ autonomy brings “higher costs, and the potential for compounding errors.”
For a first workflow, that usually means fixed steps in a fixed order, with the model doing the reading, extracting and drafting inside them. An autonomous agent can come later, once you know which cases the fixed steps do not cover.
Agent or workflow?
A workflow follows steps you define in advance. The model reads, extracts or drafts at fixed points, and the path stays the same every time. An agent decides its next step and which tools to use as it goes. That flexibility has a price: in Anthropic’s words, “higher costs, and the potential for compounding errors.”
Choose a workflow when the steps are known and repeat, such as matching an invoice against an order and a goods receipt.
Consider an agent when the path depends on what it finds, such as researching an open question across many sources, and a person reviews the result before anything leaves the building.
For a first project, a workflow is almost always the right start.
6. Launch with the team, then measure against the baseline
Launch is a training job as much as a technical one. Walk the team through the boundaries page. Show them where the system went wrong on the test set, so they know what to check. Run it alongside the old process for a short period, compare the outputs, then switch over. OpenAI’s advice fits: “Start small, validate with real users, and grow capabilities over time.”
After a few weeks, measure the same four numbers on a similar mix of cases. Report what changed, including what did not. If rework went up, fix the boundaries before you add scope.
Then pick the next workflow. Re-score the list. What you built for the first one, such as data access, an approval pattern and a trained team, often moves a neighboring candidate up.
What not to start with
The most visible demo. MIT NANDA found that budgets “favor visible, top-line functions over high-ROI back office,” while back-office deployments “often delivered faster payback periods and clearer cost reductions.” A demo for the board is not a measured workflow.
A process nobody owns. Without one accountable person, nobody sets the boundaries or acts on the numbers.
Decisions with high, irreversible risk. Credit, hiring, clinical or legal-advice decisions are poor first candidates. Start where a person reviews the output before anything leaves the building.
A workflow with no data access. If getting the documents takes a new contract or a long IT project, the work stalls while you wait.
A note on regulated work
In the EU, two parts of the AI Act matter here. This is a plain-language summary, not legal advice.
AI literacy (Article 4). This obligation has applied since 2 February 2025. The Digital Omnibus on AI, in force since 27 July 2026, rewrote it. Providers and deployers “shall take measures to support the development of AI literacy” of their staff and others who operate and use AI systems on their behalf. The obligation does not require them to guarantee any specific level of literacy. In practice: train the team at launch, and keep a short record of who learned what.
Human oversight (Article 14). For high-risk AI systems, Article 14 requires that people can effectively oversee the system. That includes the ability to “disregard, override or reverse the output” and to stop it. According to the European Commission, the high-risk rules now apply from 2 December 2027 in areas such as employment, education and critical infrastructure, and from 2 August 2028 for AI built into regulated products. Whether your workflow falls into those categories is a question for your legal team.
Either way, Article 14 makes a good design checklist. The person approving should understand the system’s limits, know the pull of what the Act calls “automation bias,” be able to override the output, and be able to stop the system. Sector rules, such as quality requirements in life sciences, professional rules in legal work and data protection everywhere, apply on top. For a broader frame, NIST’s voluntary AI Risk Management Framework organizes this work into four functions: Govern, Map, Measure and Manage.
Have a shortlist? Book a fit check. We’ll look at your top candidate and tell you whether it’s worth a diagnostic.
Frequently asked questions
What makes a good first AI use case?
A process that runs often, takes real time per case and depends on text such as documents, emails or forms. It needs a clear definition of done, errors that are cheap or caught early, data you can reach, a natural approval point and one named owner. Owner, data access and cost of error are gates.
What should not be your first AI workflow?
The most visible demo, a process nobody owns, a workflow with no data access, and decisions with high, irreversible risk such as credit, hiring or clinical decisions. Start where a person reviews the output before anything leaves the building.
What should you measure before you build?
Four numbers over two to four weeks of normal work: cycle time, hands-on time per case, rework or error rate, and backlog. Also set aside real past cases with known correct outcomes as a test set. You reuse both after launch to show what changed.
Should the first project be an AI agent or a workflow?
Almost always a workflow. A workflow follows steps you define in advance, with the model reading, extracting or drafting at fixed points. An agent picks its own next step, which brings higher costs and the risk of compounding errors. An agent can come later, once you know which cases the fixed steps miss.
Related reading
Sources
Anthropic, Building effective agents, December 19, 2024.
OpenAI, A practical guide to building agents, 2025.
McKinsey & Company, The state of AI in 2026: On the road to ROI, August 25, 2026. Online survey of 1,719 participants in 97 nations.
McKinsey & Company, The state of AI in 2025: Agents, innovation, and transformation, November 5, 2025.
MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary findings.
Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal, July 24, 2026.
Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14.
European Commission, AI Act, page updated August 3, 2026.
NIST, AI Risk Management Framework (AI RMF 1.0), January 2023.