The proposal looked healthy. The customer approved it, the team delivered and the invoice went out. Then somebody reviewed the hours and purchasing costs. The job had kept everyone busy while leaving much less money than expected.
The explanation may sit across several places: an extra revision in email, a supplier substitution, a second site visit and a delivery assumption that nobody revisited. Individually, each decision looked manageable. Together, they changed the economics of the job.
AI offers a practical way to assemble those fragments for review while there is still time to act. For a small business, the useful starting point is a regular comparison between the agreed job and the work now being requested or delivered.
A project can drift without a dramatic mistake
Consider an illustrative agency project sold for €12,000, with €8,000 in planned direct delivery costs. That leaves €4,000 before overhead and other business costs.
The client requests additional work costing €900 to deliver. A supplier change adds €600, and a further round of revisions uses another €500. With revenue unchanged, direct costs reach €10,000 and the remaining contribution falls to €2,000.
The example assumes all amounts are net of VAT and comparable on the same cost basis. It is deliberately simple, but it shows why monitoring revenue alone misses a material change. The job still looks like a €12,000 sale even though half the planned contribution has gone.
Some extra work may be a deliberate investment in the relationship. The owner should be able to see and choose that trade-off while it is happening.
Connect the evidence behind the change
Begin with the accepted quote or contract, its stated assumptions, the delivery plan and any agreed changes. Add relevant project correspondence, approved purchase records and the team’s time or cost records.
Recent connected-agent products make this kind of cross-source review a plausible workflow to evaluate. Anthropic’s Claude for Small Business announcement , for example, describes connections to business tools and includes a margin analyser among its workflows. That announcement does not validate a particular company’s accounting or establish the accuracy of its project analysis.
Your implementation needs a reliable way to identify which records belong to the job. A shared customer name may cover several projects. Match project identifiers and keep uncertain matches out of the calculation until a person resolves them.
Separate interpretation from arithmetic
AI can help identify a message that appears to request something beyond the original deliverables. It can extract the relevant passage and compare it with the accepted scope. That interpretation should remain linked to the source documents and open to review.
Calculate totals, rates and differences with ordinary spreadsheet formulas or code. Reconcile extracted numbers to the source records before using them. A fluent written explanation does not validate a figure.
Keep actual costs, committed purchases and estimates of remaining work separate. Avoid counting both a purchase order and its later invoice as two costs. Include the date of the review and show missing records explicitly.
For a job still underway, comparing revenue only with costs posted so far can give a misleadingly healthy picture. The project owner needs to review the estimated cost of completing the work as well.
Produce a short list of decisions
The output should help someone decide what to do this week. A useful review item identifies the possible change, links the evidence, explains the commercial question and names the person responsible for resolving it.
For the agency example, the assistant might flag a request for an additional format that does not appear in the accepted deliverables. The project manager then checks whether it is already covered, whether it was promised elsewhere or whether a new agreement is needed.
A supplier price increase could trigger a review of the purchase plan. Additional recorded hours might reveal an underestimated task, rework or incomplete time coding. The same variance can have several explanations; the system should not turn uncertainty into an accusation against a customer or employee.
Limit the report to items with a clear source and a possible action. Repeating every email as an alert creates another task for the project manager.
A flag is not permission to charge
A possible scope change does not automatically create a right to invoice. The responsible person must check the agreement, the commercial context and any approval needed before changing the work or price.
The workflow can prepare a comparison, an internal question or a proposed customer discussion. Keep invoices, commitments and customer messages behind the business’s approval process. Do not let an interpretation of an email silently become a new invoice line.
Also record deliberate exceptions. If the owner agrees to absorb a small addition, mark that decision so it does not return as an unresolved alert every week. The purpose is to support commercial judgment with visible evidence.
Test on jobs whose outcomes you know
Choose a few completed projects with different outcomes. Include one that went to plan, one with agreed changes and one where costs exceeded expectations. Have the project owner identify the important events before reviewing the assistant’s findings.
Check whether the workflow found those events early enough to have helped. A system that discovers an overrun only after the final invoice offers little support for the next live job.
Count false alarms and missed changes. Test revised quotes, cancelled purchases, credits, duplicate documents and work that was explicitly included in the original price. Keep some projects outside the setup work for a separate evaluation.
The owner should be able to trace every reported amount and commercial concern back to an authorised source. If the source is unavailable, the report should say what could not be checked.
Measure better decisions on active work
After a successful retrospective test, try a weekly review on a small number of live jobs. Record the time required to prepare and review it, which flags led to useful decisions and which were dismissed.
Distinguish visibility from recovered money. Finding a €500 variance does not mean the business saved €500. A benefit may come from agreeing additional paid work, changing delivery, avoiding a duplicate purchase or estimating the next project more accurately.
Keep a simple record of the action and its outcome. That gives you evidence for whether the workflow deserves ongoing attention and helps prevent exaggerated claims about its return.
Put one project under a clearer lens
XYZ by FORMATION can help connect project sources, build repeatable comparisons and put review steps around the decisions that affect customers and money. Our role is to make the workflow usable by the people responsible for delivery.
Start with a team workflow deep dive . Bring one completed project and the records that explain how it changed. We can assess whether a regular AI-assisted review would give your team a useful earlier warning on the next job.
Product description checked on 12 October 2026. The agency project and all financial figures are illustrative. They are not customer results, accounting advice or a determination of contractual entitlement.