Skip to content
Start a Project
AI & Technology

Where AI Actually Belongs in a Business Workflow

The interesting question is not whether AI can do a task. It is whether putting AI at that exact point in a process makes the process cheaper, faster or more reliable — and what happens on the day it is wrong.

Illustration generated in code — replace with a project screenshot or photograph where one exists.

A surprising number of AI projects are technically successful and operationally useless. The model works. The demo is convincing. Six months later nobody is using it, because it was placed at a point in the process where being right 90% of the time is worse than being predictable 100% of the time.

Placement is the whole problem. This is a practical way to think about it.

The wrong question

The common starting point is "what could we use AI for?" That question produces a list of tasks AI is capable of, which is now an extremely long list and therefore not useful for deciding anything.

A better starting point is a process map with time and error rates attached. Where does work sit waiting? Where does the same information get entered twice? Where does a decision depend on someone reading something and forming a judgement? Those three patterns — waiting, duplication, and judgement over unstructured input — are where AI tends to pay for itself.

Three places AI earns its place

Turning unstructured input into structured data

This is the least glamorous and most reliable use. Emails, PDFs, scanned forms, free-text notes and chat transcripts all carry information that some person is currently retyping into a system. Extraction converts that into fields.

It works because the output is verifiable. A extracted invoice number either matches the document or it does not. You can measure accuracy, set a confidence threshold, and route everything below it to a person. The failure mode is visible.

Routing and triage

When work arrives in a single queue and has to be sorted — by topic, urgency, team, or language — classification does the sorting faster than a person and with more consistency than a rota of people. The cost of being wrong is low, because a misrouted item gets bounced rather than lost.

The key design detail is that the model assigns, and a human can always reassign. The moment a misclassification becomes unrecoverable, the economics change completely.

Drafting, never sending

Replies, summaries, reports, specifications, release notes: the blank page is expensive and the edit is cheap. AI-generated drafts that a person reviews and sends remove most of that cost without moving any decision away from the person accountable for it.

The important word is drafting. The value is in the first 80%; the risk lives almost entirely in the last 20%, which is exactly the part a human reviewer is good at.

Where it does not belong

Three patterns come up repeatedly, and all three are worth resisting:

  • Deterministic tasks. If a rule can express it, write the rule. A tax calculation, a discount tier, an eligibility check — these want code, not inference. Code is testable, explainable and free to run.
  • Irreversible actions. Anything that moves money, deletes data, or sends something to a customer without review needs a confirmation step. Not because the model is unreliable, but because the recovery cost is asymmetric.
  • Decisions requiring an audit trail you cannot produce. If a regulator, a customer or a court might ask why a specific decision was made, "the model determined it" is not an answer. Either the decision is explainable or it stays with a person.
Rule of thumb

AI belongs where the output is checkable and the failure is recoverable. When either of those is missing, you are not building an AI feature — you are transferring risk.

Designing the handoff

Every useful AI integration has a boundary where the model stops and a person starts. That boundary is the part worth designing carefully, and it is usually the part that gets the least attention.

Three things make a handoff work:

  1. A confidence signal. The system needs to know when it is unsure. Without that, every output looks equally trustworthy and reviewers stop reviewing.
  2. A queue, not an interruption. Low-confidence items should collect somewhere a person checks deliberately. Scattering them into an inbox guarantees they are handled inconsistently.
  3. A visible source. Whatever the model produced, the reviewer needs the original input beside it. Reviewing an extraction without the document is not reviewing.

A checklist before you build

Before committing to an AI integration, these questions are worth answering honestly. If more than two have no good answer, the integration is probably premature.

QuestionWhy it matters
What does this cost us today, in hours or delay?No baseline means no way to tell whether it worked.
How will we know when the output is wrong?Determines whether review is possible at all.
What happens to the work if the model is unavailable?An integration with no fallback becomes a single point of failure.
Who reviews the exceptions, and when?An unowned queue is an abandoned queue.
Could a rule do this instead?Often yes, and it would be cheaper and more predictable.

Starting small, on purpose

The most reliable way to get AI into a business is to pick one process, instrument it so you know what it currently costs, integrate at a single point, and keep the human review step in place longer than feels necessary. The threshold can be relaxed once you have data; it is far harder to add review back after something has gone wrong in production.

That approach is slower than a platform-wide rollout, and it is the reason the integration is still running a year later.

Keep reading