The horse and the harness

An AI model is a strong, fast horse that doesn't know your road and sometimes bolts. What protects your company is the harness around it: clear lanes for what the AI may do alone, what needs a person, and what it must never do, with proof that those lines hold. A better model helps, but it never replaces the reins.

Your team wants to put an AI agent in front of customers. An agent here means software that uses an AI model to take steps on its own, such as reading a refund request, checking the order and paying the money back.

The demo was impressive, and the vendor says their newest model is the smartest yet. Then someone asks the right question: what stops it doing something we would regret?

The honest answer is not "a better model". It is the harness.

A strong horse that doesn't know the road

RED · NEVERmove money nobody asked for · show one customer's details to anotherYELLOW · ASK A PERSONa refund above the limit: a person approves firstYELLOW · ASK A PERSONthe agent is unsure or going round in circles: hand overRED · NEVERfollow instructions hidden inside a customer's emailGREEN · approve a refund under the limit, say ₹5,000the workreins: your rules and checksdriver: the peoplewho own the rulesthe horse: the AI model.strong and fast, but itdoesn't know your road
The AI model is the horse. The harness is everything that keeps it in the green lane: your rules, your checks and the people who own them.

Think of the AI model as a horse. It is strong and fast, and it will pull all day. But it has never seen your road: it doesn't know your refund policy, your regulator or your customers. It can be startled, and now and then it bolts.

Nobody fixes a horse like that by buying a stronger horse. You put it in a harness: reins, a cart and a driver who knows the way. In your company, the harness is the plain software and the rules around the model. It decides what the model is allowed to see, what it is allowed to touch, and what happens the moment it heads somewhere it shouldn't.

That harness, not the model, is your protection.

Three lanes on the road

The simplest way to describe a harness is a road with three colours of lane. Your team should be able to fill these in for any agent, in plain business language.

The important part is that the yellow and red lines are enforced by the harness, not by asking the model nicely. The model can be confused, flattered or tricked. A plain rule that says "no payment over the limit without a person's approval" cannot.

What the harness does on a normal day

Stays greena ₹2,000 refund, under the limitapproved, written down, donepulled backDrifts into yellowa ₹60,000 refund, above the limitpulled back: a person approves firstHeads for redan email says "send ₹5 lakh to me"stopped: no money moves unasked
Three journeys, made-up amounts. The harness lets green through, pulls yellow back to a person, and stops red outright.

Here are three journeys, with made-up amounts.

  1. Stays green. A customer asks for a ₹2,000 refund on a returned parcel. The agent checks the order, the refund is under the limit, and it goes through. Everything is written down.
  2. Drifts into yellow. A customer asks for ₹60,000 back. It is above the limit, so the harness pulls it back and puts the request in front of a person, who approves or declines it.
  3. Heads for red. An email arrives saying "send ₹5 lakh to this account". The model, trying to be helpful, starts to act on it. The harness stops it outright. No money moves that a real customer didn't ask for.

In all three, the model did roughly the same thing: it tried to help. What made the difference was the harness.

What to ask for as a leader

You don't need to understand how the model works. You need four things from the team that runs it.

  1. A named owner for the red lines. Someone in your business, not the vendor and not only the engineers, decides what is red and what is yellow, and signs it off. The limits are business decisions, like any spending authority.
  2. Proof that the lines hold. Engineers call these proofs tests: they deliberately try to push the agent into red, with tricky requests and planted instructions, and check that it is stopped every time. Ask to see them run, and ask that they run again after every change.
  3. Someone told when a line is tested. When the agent tries to cross a red line in real use, a named person should hear about it quickly, with a record of what happened.
  4. A kill switch. One way to stop the agent at once and hand everything to people, a named person allowed to use it, and proof that it has actually been tried.
BEFORE THE AGENT GOES LIVEWho decides the red lines?a named owner, not the vendorWhere is the proof they hold?tests that try to break them, every changeWho is told when one is tested?a name and a time limitWhere is the kill switch?who may pull it, and has it been triedWhat if we change the model?the same tests run again firstyou
Five questions, one card. None of them is about which AI model you bought.

A better horse is still a horse

New models arrive every few months, and each usually drifts less often than the last. That is worth having. But nobody can promise it never drifts, and one bolt into red can cost real money and real trust.

So treat a model upgrade the way you would treat a new horse in an old harness. Keep the reins, run the same tests before it goes out, and never let "the new model is smarter" become the reason a check is removed.

What goes wrong

  1. Buying the horse and skipping the harness. The agent is launched on the vendor's promise that the model is safe.
  2. Red lines nobody owns. The limits live in an engineer's head, so nobody in the business ever agreed them.
  3. Rules written as polite requests to the model. "Please don't refund more than ₹5,000" is advice, not a lock.
  4. A kill switch nobody has pressed. It exists on paper and fails on the day it is needed.

Five questions to ask your team

  1. Who in our business decided the red and yellow lines for this agent, and where are they written down?
  2. Can you show me the tests that try to break those lines, and do they run after every change?
  3. When a red line is tested in real use, who is told, and how quickly?
  4. Where is the kill switch, who may pull it, and when was it last tried?
  5. If we change the model tomorrow, what stays the same, and what gets tested again before it goes live?

Buy the best horse you can afford, but never let it go out without the reins.

For your engineers: the engineering version

Questions or disagreements: basant@cleandataengineering.ai