Short answer: start with the process that scores well on four factors: volume, error cost, data readiness and reversibility. Give each factor a score from 1 to 5 and multiply the four numbers. Pick a high result with no 1 in it. The sheet below shows the method, with five worked examples from real industries.
You can score your list of candidates in one afternoon. You will then have a clear reason for your first project that you can defend to the board. Once you have a winner, a free Automation Blueprint turns it into a written pilot plan. That is also how our AI integration work begins.
Why does the first process matter more than the model?
You can swap the model later. The process decides whether anyone trusts the result.
A first AI project faces three risks, and none of them is technical. The data sits in places nobody can reach. The output takes too long to check, so reviewers approve it blindly or stop using it. Or the process is so important that one visible mistake stops all AI work for a year.
In July 2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. It named four reasons: poor data quality, weak risk controls, rising costs and unclear business value (Gartner, 29 July 2024). This is an analyst forecast, not a measured result. Still, the list matches what we see in practice. "The model was not smart enough" is not on it.
A well-chosen first process removes most of these risks before anyone writes code. Scoring exists to make that choice.
What makes a process a good candidate for AI?
We look at four factors and score each one separately, because each one can fail on its own.
Volume. How often the process runs. High volume means saved minutes add up, and the system sees enough different cases to be tuned. Ten items a month rarely pay for a build. Several thousand documents a month usually do. Volume also pays for the review screens and integrations, which cost the most.
Error cost. What a wrong result costs if nobody spots it. A misfiled internal document costs a minute. A wrong customs value or payroll line costs money, a client or a fine. High error cost does not rule a process out. It rules out automation without a human check, and it raises the bar for the next two factors.
Data readiness. Can a system reach the inputs? Is there a history of correct outcomes to learn from? Does a named person own the exceptions? Many "AI projects" we are asked about turn out to be data-access projects. If invoices arrive as WhatsApp photos and supplier data lives in three spreadsheets, that is where the work starts.
Reversibility. A person sees the output before it has effect, or the action is cheap to undo. A suggestion waiting in a review queue is reversible. An e-mail sent to a client or a filing with a tax office is not. High reversibility lets you accept a model that is right most of the time, instead of one that must be right every time.
We leave two things out on purpose: how impressive the project would look, and how much a vendor says it will save. Neither tells you whether the project will still be running after three months.
How do you score a process? The sheet
Give each factor a score from 1 to 5 using the anchors below. Multiply the four scores. The highest possible result is 625.
| Factor | 1 | 3 | 5 |
|---|---|---|---|
| Volume | Fewer than a few dozen items a month | A few hundred a month, handled by one team | Thousands a month; several people work on it every day |
| Error cost | A wrong result is a small annoyance | A wrong result means rework or a credit note | A wrong result costs a client, a fine or a licence |
| Data readiness | Inputs on paper, in people's heads or in chats; no history | Inputs are digital but scattered; some history | Inputs in systems you can reach; a year or more of correct outcomes; a named owner |
| Reversibility | Acts directly on a client, an authority or a payment | Output is reviewed, but review is slow | A suggestion in a queue; one click to accept or fix; full audit trail |
On every row, a higher score is better. For error cost, a 5 means the process is worth the effort. It does not mean the process is safe. Two rules handle that risk:
- No 1 on data readiness or reversibility. Park the process, whatever the total says. Fix the data, or redesign the step so a person reviews the result. Then score it again.
- An error cost of 4 or 5 needs a reversibility of 4 or 5. High-stakes output goes into a review queue. It never goes straight to a client or an authority.
How to read the result. About 250 or more: a strong first project. From 100 to 250: a good second project, or a first one with a narrower scope. Below 100: leave it for now, and score it again when volume or data changes. These bands are our rule of thumb, not a statistical finding. We multiply instead of adding so that one weak factor pulls the whole score down.
Score with two or three people who do the work, not only with managers. They know the real volume and the real exceptions.
What does a worked example look like?
The table scores five processes from five industries. The scores are our judgement of a typical firm in each sector, based on the process patterns in our internal automation knowledge base. They are not measurements or client results. Your scores will differ, so run the sheet on your own processes.
| Process | Volume | Error cost | Data readiness | Reversibility | Product |
|---|---|---|---|---|---|
| Accounting: client document intake and posting suggestions | 5 | 3 | 4 | 5 | 300 |
| Construction: variation (change-order) capture from site evidence | 3 | 5 | 2 | 4 | 120 |
| Freight forwarding: trade document matching and customs data drafting | 5 | 5 | 3 | 4 | 300 |
| Dental clinic group: treatment-plan follow-up across several practice systems | 4 | 3 | 2 | 4 | 96 |
| Staffing: timesheet to payroll to invoice with discrepancy review | 5 | 4 | 3 | 4 | 240 |
The ranking matters less than the reasons behind each number. The next five sections explain them.
Which accounting process scores highest, and why?
Document intake at an accounting or bookkeeping practice with many clients. Clients send invoices, receipts and bank statements by e-mail, WhatsApp, shared folders and paper scans. Bookkeepers sort each file, rename it, attach it to the right client and period, and type the data in again. Structured e-invoice formats add more variety: KSeF in Poland, XRechnung and ZUGFeRD in Germany, Factur-X in France, FatturaPA in Italy.
Volume 5. A multi-client practice handles thousands of documents a month.
Error cost 3. A wrong account or VAT code means rework, and sometimes a corrected filing. Review usually catches it before it reaches the client or the tax office.
Data readiness 4. The ledger system already holds each client's chart of accounts and past postings. That history is the best guide for suggesting codes. The documents are within reach once the system reads the shared mailbox and upload folders. The weak spot is the steady trickle of unreadable photos.
Reversibility 5. The design is a review queue. The document sits on one side and the suggested posting on the other. One click accepts, edits or escalates it. Nothing is posted without approval.
That is why this process scores 300. It has high volume, structured inputs and checkable output, and the correct answers already sit in the ledger history. Regulation points the same way. Poland is phasing in mandatory structured e-invoicing (KSeF) during 2026 (Infor.pl, 2026). Check the exact dates for your firm with the Ministry of Finance. In Germany, every business must be able to receive structured e-invoices since 1 January 2025. Sending them becomes mandatory from 1 January 2027 for firms with prior-year turnover above EUR 800,000, and from 1 January 2028 for all firms (IHK Frankfurt, 2025).
When it goes wrong: the practice has few documents per client, or it cannot get API access to its ledger system. Score volume and data readiness honestly before you commit.
Why is construction variation capture valuable but harder?
For specialist contractors, the margin is set in the tender and lost in variations. Extra work is agreed on site, by phone or on WhatsApp. The crew does it. A week later nobody can prove who asked for it, how long it took or what price was agreed. The payment application then leaves out money the firm was owed.
Volume 3. A typical subcontractor has dozens of variations a month, not thousands. That is enough to matter, but too few to tune a model quickly.
Error cost 5. A missed variation is revenue the firm may never recover. A wrong claim starts a dispute with the main contractor.
Data readiness 2. This is the weak spot. The evidence is site photos, voice notes, chat messages and handwritten dockets, spread across many phones. Before AI can help, the firm needs one consistent way to record the evidence at the moment the work is agreed.
Reversibility 4. The system drafts a variation notice from the evidence. A project manager approves it before it goes to the client. That keeps the step reversible.
The result is 120: high value, held back by how the data is captured. The best first step here is often a capture routine on site, with automatic drafting added on top. After three months of better capture, score it again. Data readiness should rise to 3 or 4.
The sector has a lot of room to improve. The McKinsey Global Institute reported construction productivity growth of about 1% a year, against 2.8% for the world economy (McKinsey Global Institute, 2017).
What about freight forwarding and customs data?
Every shipment creates a stack of documents: commercial invoice, packing list, bill of lading or CMR, certificates. They arrive as e-mail attachments, and staff type them again into a forwarding system and a customs system. A typing error shows up later as wrong duties, delays or a fine.
Volume 5. With five to fifteen documents per shipment, volume is rarely the problem.
Error cost 5. Wrong customs values, commodity codes or parties cost real money and attract checks from the authorities.
Data readiness 3. The documents are digital, but formats differ by shipper and country. The forwarding and customs systems (CargoWise, Descartes, Magaya or national software) hold the history. Mapping document fields to system fields is the main work.
Reversibility 4. In a safe design the system drafts the declaration data and runs a check before submission. A customs clerk approves it before filing. The system never files on its own.
The result is 300, the same as accounting. Rule 2 applies here: an error cost of 5 requires reversibility of 4 or 5, and the design must deliver it. The regulatory pressure is real. CLECAT and FIATA guidance on ICS2 says forwarders acting as carriers must become house-level filers and keep control of house-level data themselves. No specific rules split filing responsibility, so the burden falls on whoever holds the data (FIATA, 2024). Correct house-level data is now worth more, and errors cost more.
Does a clinic group's central operation score well?
A dental group that has bought five to forty practices usually runs several practice management systems side by side. The central team (finance, operations, patient contact centre) works across all of them. One common leak is treatment plans that were quoted and never followed up, because nobody can see them across sites.
Volume 4. There are many open plans across the sites. Follow-up is a question of priorities, not of processing large numbers of documents.
Error cost 3. A follow-up sent at the wrong time or in the wrong words is annoying, not a disaster. Clinical and patient-communication rules still apply.
Data readiness 2. This is the structural problem. Each acquired practice brings its own system. Without a shared data layer, there is no single view of patients, plans or schedules.
Reversibility 4. The system gives a coordinator a ranked follow-up list. The coordinator decides whether to send each message.
The result is 96. It is a worthwhile project, but not a first one. For a group in this position, the honest advice is to build the shared data layer first. Other processes need it too: recalls, filling cancelled slots, claim completeness. In the UK, incomplete clinical data costs money directly. NHS dentistry pays in Units of Dental Activity, based on the clinical data entered on each FP17 claim. Since 25 November 2022, Band 2 has been split into 2a, 2b and 2c, worth 3, 5 and 7 UDAs (NHSBSA). Once the data is unified, a claim-completeness checker makes a good second project.
And staffing timesheets?
A staffing agency handles the same hours three times. It collects them from the worker or the client, turns them into payroll, and turns them into the client invoice. When the three disagree, the result is disputes, credit notes and unhappy workers. The margin per worker is thin, so billing errors hurt.
Volume 5. Every placed worker submits timesheets every pay period.
Error cost 4. Wrong pay hits workers directly and creates compliance risk. Wrong invoices cause disputes.
Data readiness 3. The ATS, payroll and finance systems exist. But timesheets arrive in many forms: client portals, e-mailed sheets, photos.
Reversibility 4. The system matches the three records and flags differences for a payroll clerk to resolve. It never releases pay or sends invoices by itself.
The result is 240. Regulation adds weight. In Germany, a worker can be placed with the same client for at most 18 months in a row. Equal pay applies after nine months (up to 15 months under a collective agreement). Exceeding the 18-month limit can cost up to EUR 30,000 in fines, and equal-pay violations up to EUR 500,000 (IHK Berlin). Timesheet data also drives those clocks, which is one more reason to build it properly.
One warning for this sector. Under the EU AI Act, AI used for recruitment, selection, candidate evaluation or allocating work is high-risk (Annex III). The cited analysis notes that most matching and ranking tools involve profiling, so the exemption rarely applies. Deployers must keep logs, ensure human oversight and inform affected workers (artificialintelligenceact.eu). Timesheet and payroll matching is back-office work. Candidate ranking is not. Score them separately, and expect extra work for the second.
How do you turn the scores into a first project?
Sort by result, then apply the two rules. Then answer three practical questions the sheet does not cover.
- Who owns the exceptions? Every automated process has cases the system cannot handle. Without a named owner for those cases, the project slowly breaks down. If you cannot name the person, the process is not ready.
- How fast can a reviewer judge the output? Output that takes seconds to check builds trust: a suggested posting, a drafted declaration. Output that takes an hour to check will not get checked.
- Does it share data work with the next project? Sometimes a lower-scoring process is the right start, because it builds the data layer or integration the next three projects need. The dental example shows this best.
Keep the first project narrow: one process, one system integration, one review screen, one audit trail and one owner on your side. Measure it against a baseline you recorded before launch. For example, track minutes per item for a few weeks first. Then "it saves time" becomes a number from your own data, not a vendor's claim. We do not quote savings percentages in this article. The honest number is the one you measure yourself.
Check the regulatory dates before you plan. The EU AI Act applies in stages. Prohibitions and AI literacy duties apply from 2 February 2025, general-purpose AI obligations from 2 August 2025, and most other provisions from 2 August 2026. The tracker we checked lists Annex III high-risk obligations at 2 December 2027 (AI Act timeline, checked 23 October 2026). Dates can still change, so verify them before you rely on them. Most back-office document processing is minimal-risk, which makes it a good first step.
What does a first project look like in practice?
Here is one real client project, anonymised. A European B2B company ran its operations by hand. We built 10 automated processes on n8n around orders, sales questions and partner onboarding. Slack is the interface, because the team already works there. The full write-up covers all ten. Figures there are adjusted for confidentiality, with ratios kept.
The project followed the same rules as the sheet above.
- Reversibility by design. Anything that involves money waits for a person to click a button in Slack. Every action is checked again after it runs, and every failure reaches a named owner.
- A narrow, high-volume first target. Customers add lines to open orders all day. Each hour before the order's set assembly time, the system compares what the customer added with what is in the order. A manager fixes a gap with one click.
- Exceptions shape the rules. One data source once lagged 15 to 40 minutes behind, which caused a false alarm. The rule since then: read the primary source again before alerting anyone.
- A baseline before launch. We tested the sales-support Slack bot on thousands of real, labelled questions before it went live. The share of questions it ignored fell from 36% to 5%.
Your project will have the same shape. First, a short discovery step to score your processes. Next, a narrow build with a review screen and an audit trail. Then a measured comparison with the baseline. After that, the next process, which reuses the integrations already in place.
FAQ
How much does a first AI automation project cost?
The model is the cheap part. Most of the cost sits in the integrations and the review screen. A pilot with one process, one system integration, one review screen and an audit trail usually takes a few weeks of senior engineering. Running costs then grow with the number of documents. We do not publish a fixed price here. Ask every supplier for three separate numbers: the build, the monthly running cost, and the time your people will spend reviewing output.
Can we use AI on customer data and stay GDPR-compliant?
Yes. Settle three points before the build. First, your legal basis and your role (controller or processor). Second, where the data is processed: EU-region hosting and a data processing agreement with the model provider. Third, how long inputs and outputs are kept. The EDPB adopted Opinion 28/2024 on personal data in AI models on 18 December 2024. Ask your data protection officer or counsel to confirm the setup for your case.
Our ERP already has an AI feature. Why build anything?
If the built-in feature covers your process from start to finish, use it. A custom build makes sense in three cases: the process crosses several systems, your rules differ from the vendor's defaults, or the work arrives as e-mails, PDFs and photos that never reach the ERP in clean form. Score the process with the sheet in this article. If the ERP feature gets the same score, you do not need a custom build.
Our first AI pilot failed. Is the process wrong or the approach?
Check the four factors in order. Most failed pilots had weak data readiness, nobody named to own the exceptions, or output that nobody could check quickly. Gartner's 2024 forecast gave poor data quality, weak risk controls, rising costs and unclear business value as the reasons projects are dropped after proof of concept. Re-score the same process honestly before you drop it. If it scored well and still failed, look at the integration and the review design. The idea is rarely the problem.
Next step
Score your top five processes with the sheet above. Then describe the winner in the free Automation Blueprint. Within a few hours you get a written plan for that process: options, an estimate and a 4-week pilot plan, reviewed by an engineer. To see how we run this work, read the AI integration service page.