What has to work before putting an AI agent into a process
An agent multiplies the process it finds, it doesn't fix it. The four things that have to be in order first, and the three you don't need to do.
Ignacio Puig Moreno
Co-founder · Business
It happens to me often: someone asks me to put an AI agent into a process that doesn't even work well with people. That's putting the cart before the horse, and it gets expensive in a fairly specific way. The process doesn't get fixed. It gets faster.
Why doesn't an agent fix a broken process?
Because it does the same thing the team does, more times and faster. If inquiries get lost today at the handoff step, they'll keep getting lost, only at a speed where nobody watches them go, and without sitting in a person's inbox where someone would eventually have noticed.
An agent clones a criterion. With no criterion, it clones improvisation and makes it consistent, which is the worst possible version: a mistake that used to happen sometimes now happens the same way every time.
What does it mean for a process to "work with people"?
It doesn't mean perfect, and it doesn't mean certified. It means two things you can check: that somebody can explain it end to end without hesitating, and that two different people on the team do it roughly the same way.
There's a twenty-minute test. Separately ask two people who do that task every day to describe how they do it. If the two versions differ on the steps, there's no process: there are two habits coexisting. Automating that means picking one of the two without having decided to, and finding out three months later.
What has to be in place before starting?
Four things, and none of them are technical.
A process owner with a name. The person who can decide how an unusual case gets resolved without asking anyone's permission. Stalled projects are rarely missing technology: they're missing somebody with the authority to answer judgment questions same-day.
The criteria written down, not the sequence. The sequence is the easy part and it's usually already in everyone's head. What needs writing is what happens when the case isn't the typical one: what counts as a properly resolved inquiry, when it goes to a person, when follow-up stops, what to say when someone asks about something the company doesn't sell.
The data in a system. If the status of an order is only knowable by asking someone who's watching a WhatsApp thread, no agent will be able to answer it. The prerequisite isn't buying a new system: it's the data being recorded in the one that already exists.
A starting number. How many inquiries arrive per day, what share gets a reply, how long the first reply takes. All countable by hand in a week. Without that, the argument about whether it worked gets settled by impression, and impressions always side with whoever talks loudest.
Before starting
How many of the four are in place?
If the open question is what you already have running, there is a tool that names it:
What you already have, in three questionsAnd what doesn't need doing first?
Three things that stall projects that were ready to start.
You don't need to change systems. Not the CRM, not the ERP, not the management software, not the spreadsheet where everything gets written down today. The agent gets built on what's already there, and a migration is an entire project that can take longer than the whole implementation. If the current system is bad, it will stay bad with or without an agent: replacing it is a separate decision, not a prerequisite.
You don't need to document the whole company. The stretch being automated is enough. A complete process manual is a months-long project that's usually out of date before it's useful.
You don't need the process to be optimal. The bar is explainable and repeatable, not perfect. If the bar is perfection, nothing ever starts.
What if the process doesn't exist yet?
Then the order is different, and cheaper than it sounds: do it by hand and seriously for a few weeks, with one person and a spreadsheet, writing down every case that doesn't fit.
Those weeks aren't lost time, they're the discovery phase. Finding an exception with a person costs an awkward conversation. Finding it with an agent already handling real traffic costs a customer.
When does optimization happen?
After the ordering, and it's the part almost nobody uses. The most common request once a process is written down is for the agent to do exactly what the person did, step by step.
And a good share of those steps exist because a human needs to remember something, wait for somebody else to reply, or hand a piece of data from one system to another. None of that is needed when whoever executes checks both systems at once and doesn't forget.
Illustrative example
The steps that disappear
12 steps
- 01
Notice that a new message came in
the executor asks for it - 02
Read the inquiry and understand what it asks for
the process asks for it - 03
Ask what they are looking for, for what space and in what area
the process asks for it - 04
Write the answers on paper or hold them in memory
the executor asks for it - 05
Check in the system whether that person was already a customer
the process asks for it - 06
Copy the conversation data into the system
the executor asks for it - 07
Decide whether the inquiry qualifies and at what priority
the process asks for it - 08
Find out who is available in that area today
the process asks for it - 09
Forward the conversation to the right salesperson
the executor asks for it - 10
Remember to follow up 48 business hours later
the executor asks for it - 11
Write the follow-up message
the process asks for it - 12
Mark the inquiry as cold if they never replied
the executor asks for it
Twelve steps, and none of them is optional when on the other side there is one person alone with a phone in hand.
An illustrative example of a typical commercial process, not a measurement from a customer. The point is not the number: it is that half the steps of almost any manual process exist to compensate for the limits of whoever executes it.
Ordering first and optimizing after isn't bureaucracy: it's the only thing that shows which steps belonged to the process and which belonged to the human body executing it.
An ordered process without an agent keeps working, slower. An agent on top of a disordered process multiplies the disorder, and does it at a speed where nobody can watch it anymore.
Frequently asked questions
How long does it take to order a process before automating it?
It depends on the stretch, not on company size. For a specific task, like qualifying and routing incoming inquiries, two or three one-hour sessions with the person who does it daily plus a week of reading real conversations is usually enough. If the ordering turns into a months-long project, the scope was cut wrong.
Do you need particular software in place before starting?
Not a new one. What's needed is for the data the process depends on to live in a system that can be queried from outside, and to find out early whether that system allows it, with what permissions and at what cost, because it varies a lot between vendors.
Should the same vendor that implements the agent also do the process work?
It's usually fastest, because the uncomfortable questions surface while mapping. What matters is that the stage has its own deliverable: the written process has to end up in the company's hands and be useful even if the agent project doesn't continue.
Ignacio Puig Moreno · Co-founder · Business, StudioChat
Want agents like these working for your team?
Talk to us now