To pick a first AI automation project, score each candidate process from 1 to 5 on volume, rule clarity, data access and risk (5 is always the better end), weight the four scores, and rank the totals. Then apply three gates: a named owner, a measurable baseline and the right to use the data. The highest total that passes the gates is your first project.
The first AI automation project has an outsized job. It has to work well enough to earn trust, and it has to be measured well enough to justify the second one. Teams usually choose it by enthusiasm: the process the loudest manager complains about, or the one that looks best in a demo. A scoring matrix replaces that with a short, written comparison that you can show a sceptical finance director. This guide gives the matrix, explains how to score each criterion with evidence, works through an example with made-up numbers, and shows what to do when the winner turns out not to need AI at all. The wider picture is on our AI automation for business page.
Why a scoring matrix beats a shortlist by feel
A matrix does three useful things. It forces each candidate to be described in the same terms, so a quiet but high-volume process can beat a noisy low-volume one. It exposes disagreements early: if two people give the same process different rule-clarity scores, they do not share an understanding of the process, and that is better learned before the build than during it. And it leaves a record, so when the first project succeeds or stalls you can see which assumptions held.
It is not a formula that makes the decision for you. It is a structured way to argue.
The four scored criteria
Each criterion is scored from 1 to 5, and 5 is always the better end. That means "risk" is scored as safety: a process where a mistake is cheap and caught quickly gets a 5.
| Criterion | Question | Score 1 | Score 3 | Score 5 |
|---|
| Volume | How many items does the process handle each month? | A few dozen | Hundreds | Thousands, or a few hundred that each take long |
| Rule clarity | Could two experienced people agree on the correct result for most items? | They would often disagree | Agree on most, with a grey area | Agree on nearly all, and the rules can be written down |
| Data access | Can the system reach the documents, records and systems it needs, lawfully? | Data is scattered, on paper or restricted | Reachable with some preparation | Already digital, in systems with interfaces, cleared for this use |
| Risk (scored as safety) | What happens if an item goes wrong and is not caught? | Money, safety or a regulator is affected | Rework and an unhappy customer | Cheap to catch and fix, internal only |
Volume is about repayment, not importance. A task done ten times a month will not repay an automation even if it is annoying. Use counts from systems, such as tickets, invoices, emails, orders or forms, not estimates from memory.
Rule clarity predicts whether you can write down what "correct" means, which is what an evaluation set requires. If experts disagree, the model has nothing stable to learn from or be measured against.
Data access is the most common reason pilots stall. Count the real steps: where the documents live, who can export them, how the system would be given access, and whether processing them this way is allowed. Personal data and customer data need an explicit decision about where they may go, covered in how to evaluate an AI pilot.
Risk should consider the worst plausible miss, not the average one. A process that is usually harmless but occasionally touches payments deserves a 2.
Weights express what matters to your business. A reasonable starting set:
- Volume 30%
- Rule clarity 25%
- Data access 25%
- Risk (safety) 20%
Total score = 0.30 × Volume + 0.25 × Rule clarity + 0.25 × Data access + 0.20 × Risk. The result runs from 1 to 5; multiply by 20 if you prefer a number out of 100.
Volume carries the most weight because without it there is no return. Risk carries the least in the default because you control it with review gates, but a regulated or safety-critical business should raise it.
Three gates that scoring cannot replace
Apply these after scoring. A process that fails a gate is not the first project, whatever its total.
- A named owner. One person in the business accepts or rejects the result and has authority to change the process around it. Without that person, nobody can say the pilot worked.
- A measurable baseline. You can state today’s volume, time per item and error or rework rate. If you cannot, spend a few weeks measuring before building; otherwise the return can never be shown.
- The right to use the data. Someone with authority has confirmed the data may be processed this way and by whom. In practice this means your security or data-protection lead has seen the data path.
A worked example (illustrative numbers)
The example below is invented to show the method. It is not a client, a benchmark or a result. Imagine a distributor with several hundred staff that lists six candidates. Scores are written as Volume / Rule clarity / Data access / Risk (safety).
| Candidate process | V | R | D | S | Total (default weights) |
|---|
| Supplier invoice intake | 5 | 4 | 4 | 4 | 4.30 |
| Month-end bank reconciliation | 4 | 5 | 5 | 2 | 4.10 |
| HR policy questions from staff | 3 | 4 | 5 | 4 | 3.95 |
| Customer order emails | 4 | 3 | 4 | 3 | 3.55 |
| Sales call summaries | 4 | 2 | 3 | 4 | 3.25 |
| Contract clause review | 2 | 2 | 3 | 2 | 2.25 |
Check one row by hand: supplier invoice intake = 0.30 × 5 + 0.25 × 4 + 0.25 × 4 + 0.20 × 4 = 1.50 + 1.00 + 1.00 + 0.80 = 4.30.
Read the table in three passes.
Pass one: the ranking. Invoice intake leads on volume and has fairly clear rules, since an invoice either matches a purchase order or it does not. Contract review is last: low volume, experts who would disagree, and a high cost of error.
Pass two: does the winner need AI? Month-end bank reconciliation scores second, but its rules are exact and the data is structured, so plain workflow rules are the right tool, not a model. The matrix tells you where automation pays; it does not tell you which kind. Run each top candidate through the AI agents vs RPA vs workflow automation decision table. Invoice intake starts with PDFs in many layouts, so it needs an AI step with rules around it. Reconciliation goes to the workflow backlog and may not need AI at all. For reconciliation, how bank data reaches the ledger also matters; see bank feeds vs nightly statement imports for reconciliation.
Pass three: does the ranking survive a change of weights? Suppose the finance director insists that risk must count for 35%, with volume 25%, rule clarity 20% and data access 20%. Recomputed, the totals become: invoice intake 4.25, HR policy questions 3.95, bank reconciliation 3.70, customer order emails 3.45, sales call summaries 3.40 and contract review 2.20. Invoice intake still leads; bank reconciliation drops below HR policy questions. A winner that holds under different weights is a robust choice. If the winner changes with every set, the top candidates are close, and the gates and the pilot design should decide.
The gates. Say the distributor has an accounts payable lead who owns the invoice process, counted 2,400 invoices last month with an average handling time from a sample of timed items, and confirmed with its IT lead that invoices may be processed under its existing supplier-data rules. All three gates pass, so invoice intake becomes the first project, with HR policy questions as the second, because it reuses the same skills with different data.
Collecting evidence for each score
Scores are only as good as the facts behind them. A short evidence pass takes a day or two:
- Volume: export counts for the last three months from the system of record, and note seasonal peaks.
- Rule clarity: ask two people who do the work to process the same twenty items separately, and compare. The share they agree on is a direct measure. This also becomes the seed of your evaluation set.
- Data access: list each source, the owner, how it can be exported and whether it contains personal data.
- Risk: write down the single worst miss in the last year and what it cost.
- Baseline: time a sample of items with a stopwatch, including the exceptions. Do not estimate.
What to do with the losers
Keep the full table. The candidates that fail on data access are your data-preparation backlog. The ones that fail on rule clarity are a process-design backlog, where writing down the rule often solves much of the problem before any AI is involved. Those that fail on risk are candidates for later, once review gates and logging are proven on a safer process.
Turning the winner into a pilot
With the first project chosen, write a one-page pilot brief: the process and its owner, the baseline, the pass and fail thresholds, the data used, who reviews uncertain items and what happens at the end. Then estimate the return with the AI automation ROI formula and worked example, because a project that scores well but cannot cover its running costs is still the wrong first project. Once a pilot has worked, scaling AI pilots to department-wide workflows covers why pilots stall and what to check before scaling.
Timeline Digital is a custom software development company founded in 2013 and headquartered in Islamabad, Pakistan. This guide was written by Usama Asif. If you would like help scoring your own candidates, bring the list and a sample of each process to a conversation through the contact page. There is no fixed timeframe: the schedule is set from your requirements in the Plan step. The free pilot of 2 to 3 key modules is ready in about 2 days, and with 1,200+ developers we have the capacity to deliver a complete system in as little as 2 weeks. The full project starts after you approve the pilot, and the scope, milestones and terms are agreed after it. Long-term support under a service level agreement (SLA) follows launch.