The last step of an AI workflow is where the time savings quietly disappear. Here is what our partners taught us, and what we changed.
A partner told me they were happy with the requirements document our platform produced. Good structure. Nothing missed. Hours of work done in minutes.
Then I asked what they did with it next.
They opened their own spreadsheet and copied every line across by hand.
I want to explain why that happened, why it took us a while to see how big a problem it was, and what we did about it. Because I think it points to something bigger about AI at work, not just in ERP.
Imagine you order something heavy. A courier drives it four hundred miles, gets it to your street, then parks and says the rest is up to you. You carry the box the last hundred yards yourself.
Nothing about that is a scam. You still saved four hundred miles. But you remember the last hundred yards, because that's the part that cost you the effort.
Most AI tools at work are that courier. They do the heavy lifting and then hand you something you still have to fix up before anyone else can see it. Reformat it. Restructure it. Put your logo on it.
That final stretch does not look like much on a plan. In practice it is where the promised time savings go to die.
Quick context, because the rest of this will not make sense without it.
Infomind runs ERP rollouts. Four AI agents do the heavy, repetitive work that normally eats months of a project:
These agents do not summarize. They produce the real deliverables, the documents and scripts a project actually runs on.
And for our first stretch as a company, they all produced those deliverables in one layout. Ours.
At the start, that was fine. We had a clean design, it was clear, and early customers were happy to use it.
Then more customers and more delivery partners came on board, and the same request kept arriving from all of them. Not a complaint about quality. A request about shape.
A system integrator like Columbus or Wipro does not run one project. It runs many, for many different clients, over many years.
In that time it builds its own standards for how a requirements document should look. Which columns appear, in what order, what each one is called, what has to be filled in before a requirement counts as done.
Those standards are not a preference. They are what years of delivery experience taught them. Most of those columns exist because a project once went wrong without one, and somebody made sure it would never happen again.
Our default format was never going to reflect that, and it never could. We have not sat in their delivery reviews. We did not learn their lessons. Asking a partner to drop their standard and use ours was really asking them to leave their own experience at the door.
The same is true on the customer side. A manufacturer has a tracker its own business people already know how to read.
So when our platform handed over a document in Infomind's layout, the receiving team had two options. Ask everyone to learn a new format for this one project. Or retype the content into the format everybody already uses.
Almost everyone chose retyping.
That is the courier stopping at the end of the street. We had built agents to remove manual work from ERP rollouts, and the very last step of the process was manual work.
Columbus, one of our delivery partners, needed requirements written as user stories.
This was not a small change. Their format has its own columns: role, business value, expected result and acceptance criteria, MoSCoW priority, workstream, session reference. It reads as a sentence from the user's point of view. "As a Customer Service Representative, I want to receive and process customer orders from both email and EDI channels within a single ERP interface, so that order entry is consolidated and manual handling is reduced."
There was no sensible way to stretch our layout into that shape. And Columbus was not an unusual case. Columbus was simply the clearest version of what every partner and customer had been telling us for months.
One layout cannot serve everybody. Not across different integrators, not across their different clients, not across a rollout that runs for two years.
Custom Prompts lets you give the platform your format instead of accepting ours. There are two ways to do it, and neither one needs a developer.
Upload the file you already use. Point the platform at your existing Excel tracker. It reads your header row, works out your columns, and shows you the list it found so you can confirm nothing is missing.
Or just describe it in words. Type what you want in plain language. No template needed.
From that point on, the agent produces work in your structure. Your columns, in your order, with your names for them. Extra fields we never thought of. And when you download the file, it carries your logo and your styled headers, not ours.
If you have not set up a Custom Prompt, you still get the Infomind layout, exactly as before. Nothing was taken away from anyone.
Saving the retyping is the obvious win. It is not the biggest one.
The biggest one is consistency.
Think about a system integrator running fifteen ERP projects at once, with different consultants on each. Every one of those projects produces requirements documents. Whether they all look the same has always depended on the discipline of whoever assembled the document that week. Some consultants follow the house standard closely. Some drift. New joiners copy whatever file they were handed first.
Set up a Custom Prompt once, and every project produces work in the same structure with the same branding. Consistency stops being something you chase in review meetings and becomes something the platform does on its own.
For a partner whose reputation travels through the documents it hands to clients, that is worth more than the hours saved.
There is a version of AI at work that produces something impressive, then leaves you to make it usable. It looks like a finished job in a demo. It is not a finished job at your desk.
Finished work is work you can send to your client without touching it.
Anything short of that is a document you have to handle twice. And handling it twice is where AI time savings quietly go to die. Not dramatically. In an afternoon nobody logs on a timesheet.
The better AI tools are all moving the same way. Let the user supply the format, instead of dictating it to them. We are not the first to work this out, and it will not stay a differentiator for long. It will just become what people expect.
We got this wrong for a while. We built four agents that did brilliant work and then asked everyone to accept our layout. The fix was not a better layout. It was letting go of the idea that we should choose the layout at all.
The short film below shows the whole flow. Upload a template, watch the columns get detected, generate requirements as user stories, download the branded file.
Custom Prompts is available now in Infomind.AI.
Our AI created the documents. Our customers transformed how they used them. Discover the unexpected ways AI is...
The real question isn't whether AI will replace ERP consultants—it's who will thrive in an AI-powered future. Find out...
Not all ERPs are built for AI. Discover how to evaluate AI readiness and choose an ERP platform that supports long-term...