How Custom AI Solutions Drive Business Transformation

Your team already licenses an AI tool, yet one workflow, such as contract renewals or ticket triage, is still done by hand. Does it need a custom AI solution, or would the tool cover it once configured properly? This guide gives you a nine-point fit test for that one workflow, the four outcomes it leads to and a worked example.

Answer the question with a trial on real records, not a feature comparison. Use the tool if it passes. Configure it or integrate the missing piece if it nearly passes. Commission a bounded custom workflow only when the trial shows a gap that configuration cannot close. And if nobody owns the outcome or the data, settle that first.

Describe one decision, not "AI for the business"

A fit test needs a unit small enough to check. Write down one operational decision in six fields:

• Trigger: what starts the work (a renewal date, a new ticket, an uploaded contract).
• Inputs: which records and documents a person reads today.
• Authoritative source: which system wins when two sources disagree.
• Proposed action: what the AI step should produce (a draft, a recommendation, a routed case).
• Reviewer: the named person who accepts or rejects it.
• Destination: where the approved result is written, and by whom.

If you cannot name the reviewer or the authoritative source, no tool, custom or ready-made, will make the workflow reliable. Settle that with the workflow owner before any trial; no developer can decide it for you.

Where generic tools fall short — and how to check it

The usual objections to a licensed tool are missing domain rules, no connection to internal systems, weak admin controls and no audit trail. Check each one in the trial before you accept it.

Microsoft Copilot Studio, for example, documents prebuilt and custom connectors, including connectors to non-Microsoft systems, and data policies that let admins control which connectors an agent may use together. Sensitive data alone does not prove you need to build. What matters is whether the configured tool meets your workflow's requirements when it runs on your own records.

The fit-test worksheet

Run a short trial of the tool you already have, or the strongest candidate, against the decision above. For each point, record the evidence from the trial, the gap and who owns closing it.

1. Source access and freshness: can it read the authoritative source directly, and how soon do changes appear?
2. User-level permissions: does each user see only what they could already open in the source system?
3. Connector or action: is the read or write you need supported, or does it need a custom connector?
4. Approval before consequential writes: can a named reviewer approve before anything changes in the destination system?
5. Auditability: can you reconstruct which sources, version and reviewer produced a given result?
6. Failure path: on a timeout, a missing source or conflicting evidence, does it escalate or guess?
7. Task quality: how does it do on your own test cases, including hard and forbidden ones?
8. Configuration, export and exit: can you export configuration and logs, and leave without rebuilding the process?
9. Support ownership: who fixes it when it breaks: the vendor, your admins or an engineering partner?

There is no score to add up. One failed point that blocks the workflow, such as permissions or approval, outweighs several comfortable passes.

Build the task-quality test set before the trial. OpenAI's evaluation best practices recommend task-specific evaluations on cases that represent production, a clearly stated objective, and human judgement where automated scoring falls short. One impressive demo is not that evidence; responses vary, so repeat the important cases.

Four outcomes of the test

The worksheet leads to one of four decisions:

1. Use the tool as it is. Every blocking point passes. Spend the budget on adoption, source quality and the review routine.
2. Configure it. The gaps are settings, instructions, connectors or data policies. An internal admin or a short engagement closes them.
3. Integrate the missing boundary. The tool handles the language work, but permissions, an approval step or a write into your system of record needs engineering. Build only that layer and keep business rules in code.
4. Build a bounded custom workflow. Several blocking points fail for structural reasons: the logic is specific to your domain, the sources cannot be reached, or the approval and audit trail must live inside your product.

What you commission when you build

A custom AI solution rarely means training a model from scratch. You pay for the layer the worksheet showed missing: permission-aware retrieval, the rules that flag conflicts, the approval step and the audit log, on top of an existing model. Specify the data source, decision rules, reviewer and handoff first; choose the model last.

If your product already runs, use this guide to scoping the first AI-assisted workflow to define authorised sources, review and acceptance before commissioning an integration.

A worked example: a supplier renewal proposal

Consider a hypothetical procurement team. It illustrates the method; it is not a customer result.

Ninety days before a supplier contract ends, a buyer drafts a renewal proposal. Today that means reading the contract terms, checking spend against the agreed volume, pulling delivery and quality incidents from the operations system and writing a recommendation for the category manager.

• Decision: renew, renegotiate or retender, with the evidence attached.
• Sources: the contract repository (authoritative for terms), the finance system for spend, the operations system for delivery records.
• Proposed action: a draft recommendation that cites each source and flags conflicts, such as spend above the contracted volume or late deliveries the contract penalises.
• Reviewer: the category manager approves, edits or rejects.
• Destination: only after approval is the decision written to the contract record.

Now run the worksheet. Suppose the licensed assistant reads the contract repository and finance exports well, but cannot respect per-category permissions in the operations system and has no approval step before writing to the contract record. That is outcome 3: keep the tool for drafting, and integrate permission-aware retrieval and the approval-gated write. If instead the conflict rules are specific to your contracts and the tool cannot apply them reliably on your test cases, the evidence points to outcome 4: a bounded workflow with those rules in code.

Decide between one agent, several or a single prompt after the test, once you know which layer is missing.

Compare costs on the same workload

Set a baseline for cycle time and review quality before deciding what to automate. Then compare the options on the same workload and the same quality requirement:

• current time per case, including review and rework;
• licence or API cost at the expected volume;
• one-off configuration, integration or build effort;
• monitoring, evaluation upkeep and maintenance;
• the cost of leaving: exporting data and moving the process elsewhere.

A faster draft that takes longer to check has not saved anything.

When the workflow touches legal or candidate data

In LegalTech and HRTech, settle two things before the trial: which personal data may reach the model at all, and who may see each result. Point 5 then has to hold for every result, not just a sample. Adding either after delivery means rebuilding the retrieval and approval layer. In the renewal example, that means the per-category permission filter and a record of every approval the category manager gives.

What to bring to a commissioning conversation

Bring one workflow written up in the six fields above, an authorised sample of real cases, the tool you already use and the worksheet points it failed. That is enough to decide whether you need a short configuration, a bounded AI workflow integration in the product you run, AI workflow automation development for outcome 4, or an AI discovery sprint because feasibility or the investment case is still unclear.

Use the "Plan your AI roadmap" button to send us that write-up. In a free 30-minute discovery call we go through which of the four outcomes your evidence points to and what the next piece of work would be. Commission only the part the ready-made tool failed.