Do you have to clean up all company data before deploying an AI agent?
The exact minute an agent answers wrong because of data written from two places, and what actually has to be sorted first. Not the whole company.
Iván Itzcovich
Co-founder
Do you have to clean up all company data before deploying an agent? The short answer is no, and the long one is easier to see by looking at the exact minute an agent answers wrong.
Where exactly does it break?
One order, an ordinary day. At 09:10 the customer buys and it gets recorded in the management system. At 11:40 production posts in a WhatsApp group that it's delayed by two days, and the rep notes it in the spreadsheet they check every day. At 12:05 the customer asks when it will arrive, the agent queries the management system, the only place it's connected to, and answers that it ships today.
That answer is wrong, well written and confident. At 17:30 the customer writes again, annoyed, and only then does a person step in.
The problem wasn't the agent. It was the twenty minutes in which the real status lived in one place and the status the agent reads lived in another.
Illustrative example
Two systems writing the same field
The customer confirms the purchase. It gets recorded in the management system and the rep also notes it in their spreadsheet.
Management system
Confirmed
The team's spreadsheet
Confirmed
What the agent answers if the customer asks right now
Your order is confirmed. I will let you know when it leaves the warehouse.
Both places say the same thing, so the agent gets it right.
What happened
The agent never failed technically. It read the value available to it and answered in well written prose. The failure was twenty minutes earlier, when the real status got written somewhere the agent is not connected to, and nobody had any way to notice until the complaint.
Both modes, moment by moment
Two places write the status
- 09:10. The customer confirms the purchase. It gets recorded in the management system and the rep also notes it in their spreadsheet. Management system: Confirmed. The team's spreadsheet: Confirmed. What the agent answers if the customer asks right now: Your order is confirmed. I will let you know when it leaves the warehouse. Both places say the same thing, so the agent gets it right.
- 11:40. Production posts in a WhatsApp group that the order is delayed by two days. The rep notes it in their spreadsheet, which is what they check every day. Management system: Confirmed. The team's spreadsheet: Delayed, ships Thursday. This is where they start to differ. The system kept the old value and nobody updated it.
- 12:05. The customer asks on WhatsApp when the order will arrive. The agent queries the management system, the only place it is connected to. Management system: Confirmed. The team's spreadsheet: Delayed, ships Thursday. What the agent answers if the customer asks right now: Your order is confirmed and ships from the warehouse today. It should reach you tomorrow. This is the moment. The answer is wrong, well written and confident.
- 12:06. The agent logs the conversation as resolved and escalates to nobody, because from where it looks nothing unusual happened. Management system: Confirmed. The team's spreadsheet: Delayed, ships Thursday. Nobody on the team finds out. No alert fires, because there was no technical error.
- 17:30. The customer writes again because the order never arrived. Now a person does step in, and has to explain and apologise. Management system: Confirmed. The team's spreadsheet: Delayed, ships Thursday. The cost was not the error: it was the complaint, and that the customer found it before the company did.
- Next day. Somebody updates the system by hand so it matches the spreadsheet. The case gets closed and the cause is exactly what it was yesterday. Management system: Delayed, ships Thursday. The team's spreadsheet: Delayed, ships Thursday. Tomorrow it happens again with another order, because both places still write the same field.
What happened. The agent never failed technically. It read the value available to it and answered in well written prose. The failure was twenty minutes earlier, when the real status got written somewhere the agent is not connected to, and nobody had any way to notice until the complaint.
One write path
- 09:10. The customer confirms the purchase. It gets recorded in the management system, which is now the only place the status gets updated. Management system: Confirmed. Second source: No longer exists for this field. What the agent answers if the customer asks right now: Your order is confirmed. I will let you know when it leaves the warehouse. There is a single source, so there is nothing to reconcile.
- 11:40. Production reports the two day delay and enters it in the system, which is where the status is recorded now. Management system: Delayed, ships Thursday. Second source: No longer exists for this field. The team still reads the information where it always did, but writes it in one place only.
- 12:05. The customer asks when the order will arrive. The agent queries the system. Management system: Delayed, ships Thursday. Second source: No longer exists for this field. What the agent answers if the customer asks right now: Your order was delayed and ships on Thursday. I will let you know as soon as it leaves the warehouse. Same question, same hour, correct answer. The only thing that changed is where the status gets written.
- 12:05, variant. If production had not entered anything yet, the system would not hold the new value. A well built agent says so instead of improvising. Management system: Confirmed, nothing new since 09:10. Second source: No longer exists for this field. What the agent answers if the customer asks right now: I do not have a confirmed new date. I will check with the team and get back to you today. An agent that never says it does not know is badly built, not badly fed.
What changed. No company systems were unified and nothing was migrated. Somebody decided which single place updates one field, the order status, which is the one the agent needs. Every other system stayed exactly as it was.
The example is illustrative and uses an order because it is the most recognisable case, not the measurement of a customer. The same pattern shows up with a balance, an available slot, a stock level or the status of an application.
Why does nobody notice before the complaint?
Because there was no technical error. The agent read valid data from a working system, so no alert fires, no log turns red, and the conversation gets logged as resolved.
This is the difference between messy data and contradictory data. Mess announces itself: a field is missing, a query returns nothing, somebody asks. Contradiction doesn't, because both versions look equally valid from the inside.
One part of this is solved by design rather than by data: an agent has to be able to say it doesn't have the information and hand off. An agent that never says it doesn't know is badly built, not badly fed.
So what actually has to be sorted first?
One write path for every piece of data the agent will write, or will read in order to decide something. Nothing beyond that.
In the example above it's enough to decide that an order's status gets updated in a single place. No unifying systems, no migrating history, no switching software: just choosing where that field gets written and having everything else read from there.
And the rest of the company's systems?
They can stay exactly as they are. If the ERP and the CRM spell the same customer's name differently, that's a real reporting problem and it isn't a problem for the agent qualifying new inquiries, because that agent never touches the ERP.
The useful question isn't whether company data is in order, because the honest answer at any company older than five years is no. It's which data this task touches, and where it gets written today.
When does consolidation genuinely come first?
When the problem being solved is reporting, traceability or financial close, rather than a conversational task. There, consolidation is the project, not a prerequisite for a different one.
And if the company genuinely needs to consolidate, an agent doesn't replace that. What it does is produce the real map of one stretch, with measured usage, before the big budget gets spent. It's the order that risks the least money, not a way to avoid the work.
What does each path cost?
Consolidation is measured in months, involves every department and delivers value at the end. A conversational task is measured in weeks, involves the people in that stretch and delivers value when it ships. At Takenos, the support agent resolves 7 out of every 10 conversations with no human involved, at roughly a minute per ticket, by checking the real state of an account.
| Consolidating systems | Giving an agent one task | |
|---|---|---|
| Scope | Every system in the company | The data one task touches |
| Duration | Months | Weeks |
| Who takes part | Every department | The people in that stretch |
| What gets touched | ERP, CRM, management software, history | One field and its write path |
| When value shows up | At the end | When it ships |
Confusing the two projects is what turns an eight-week implementation into a year-long one, and it usually ends with neither one approved.
Frequently asked questions
Can an agent run on top of a messy CRM?
Yes, as long as the mess isn't in the fields the agent reads or writes. Look at the health of those specific fields rather than the system overall, which is rarely good and rarely the problem.
What does one write path mean?
That each piece of data has a single place where it gets created and updated, and everything else reads from there. When two systems can write the same field, somebody has to decide which one wins every time they disagree, and that reconciliation work repeats forever.
How do I know if I have this problem without reviewing everything?
By looking at one thing only: for each piece of data the agent will use, how many places update it today. If the answer is one, there's nothing to do. If it's two, that's the prep work, and it takes days.
Iván Itzcovich · Co-founder, StudioChat
Want agents like these working for your team?
Talk to us now