How do you implement an AI agent in a B2B company? start with the process, not the tool
Before the first stage of a rollout begins, one decision settles the outcome: which process you hand over to an AI agent, and whether the company is ready to hand it over. Below are the three conditions the process has to meet, a readiness test in five questions, the list of decisions nobody will take for you, and the signals that an implementation partner is promising too much.
- 3 conditionsa process suits an AI agent when it repeats, when the data sits in a system, and when the effect can be read off a number
- 5 questionsenough to judge whether the company is ready; one “no” clears in a week, two mean you sort the data out first
- 1 persona process owner with the right to decide; without one the project stops at the first question about an exception
- EUR 3,500net per week for the process workshop where the choice of process is settled; the rollout is quoted after it
Short answer
How do you implement an AI agent in a B2B company?
Start by picking one process, not by picking a tool. A process suits an AI agent when it meets three conditions at once: it repeats at least a few times a week, the data it needs sits in a system you can read it out of, and the effect can be written as a number measured before the start and again after 90 days. If one of the three is missing, pick another process, or cut out the slice of this one that does meet them.
Then run the five questions from this page and name one person with the right to decide the scope of that process. That person settles the exceptions and says “from Monday we do this differently”. Without them the project stops at the first question nobody can answer, and comes back a month later.
Only then do the stages start: workshop, pilot on real data, production, measurement. We lay them out week by week in process automation step by step. The choice of process itself is settled at the process workshop: EUR 3,500 net per week, ending with a process map, metrics and a 30-day plan you can execute without us.
Which processes suit an AI agent, and which do not?
Three conditions, and all three at once. Repetition, because an AI agent pays back on the number of repeats, not on how hard the task is. Data in a system, because the agent answers from what it can read, not from what people remember. A measurable effect, because with no “before” number, three months later you are left exchanging opinions instead of reading a report.
Below are six processes that come up most often in our conversations, and what settles the verdict on each. The verdict is not a judgement on how valuable the process is, only on how ready it is today. A process in the “no” column can be a good candidate six months later, once the data lands in one place or once somebody writes the rules down. An agent like BOND, which answers from company data in 5 to 10 seconds, is only as good as the data it can reach.
| Process | What settles it | Verdict |
|---|---|---|
| Questions about order and payment status | asked every day, the data sits in the sales and accounting systems | yes, the typical first pilot |
| A weekly report pulled from several sources | repeats every week, the sources have an API or an export | yes |
| Retyping data from PDFs into a system | repeatable, measurable in hours, the result is easy to check | yes |
| Searching the team's documents and notes | the data exists, but it is scattered across drives and inboxes | only after a clean-up |
| Setting a price for a new client | the rules are unwritten and change every quarter | no, write the rules down first |
| Negotiating terms with a key account | happens rarely, the context sits in the relationship, not in the data | no |
The “verdict” column is our own read on how ready a process is for a first pilot, not a ranking of how much it matters to the business. A process marked “no” is still worth mapping at the workshop, because there is usually a repeatable slice to cut out of it: gathering the data for a quote, yes; deciding the discount, no.
How do you check whether your company is ready for an AI agent?
The test takes fifteen minutes and needs nothing but honesty. Answer “yes” only when you can point to a specific system, a specific person or a specific number. “Probably yes” counts as a no, because it surfaces in week three of the project anyway.
Five times “yes” means you can start with the workshop and a pilot. One “no” can usually be cleared in a week, most often by naming the process owner. Two or more mean you sort out the data first, or settle who decides, and push the rollout back by a month.
- Can you name one process your team repeats at least a few times a week?
- Does the data that process needs sit in a system you can read it out of, rather than in two people's heads?
- Is there one person who can decide that from Monday this process works differently?
- Can someone grant read-only access to 2 or 3 systems within a week?
- Can you give the “before” number: hours, number of repeats, or time from question to answer?
Choosing the first process in four steps
This is not a brainstorm. It is four things to do before anyone shows you a tool. They take about a week, and you can do them yourself, with no partner and no budget. What comes out is one sheet of paper you bring to the first conversation.
Write down the questions and the tasks
For three days, note the questions that come up in the team (“how many do we have”, “where is it”, “has the client paid”) and the tasks repeated every day. Aim for a list of twenty entries, not the five most interesting ones. The dullest lines on that list are usually the best candidates.
Measure three of them for a week
Do not ask the team how long it takes. Measure it. Declarations leave out hunting for the file, fixing it, and asking a colleague which version is current. Write down the hours, the number of repeats per week and the number of mistakes somebody had to correct.
Check where the data sits
For each of the three tasks, name the system and the person who can grant read-only access to it. If the answer is “it is in Kate's inbox”, the task drops off the candidate list and moves onto the clean-up list.
Set the “before” number and the walk-away threshold
Write down one number you will measure again after 30 days and after 90 days, and the threshold below which you will call the rollout a failure. A threshold written before the start protects you from bending the verdict to fit the result, which is the most common reason a pilot never gets to end.
Why does a rollout not move without a decision maker?
An AI agent does not take work away, it changes how the work gets done. Every change like that has a few moments where somebody has to say “yes” or “we do it differently”. These are not technical decisions, and nobody from outside will take them for you, because they concern your process and your data.
That is why we ask about the process owner on the first call, before we ask about systems. A two-person company with a clear process owner rolls out faster than a company with an IT department and a steering committee, because questions about exceptions come up daily and a committee meets once a fortnight.
- Which exception in the process is actually the rule, and which happens once a quarter and stays with a person
- Whether the team stops doing the task by hand, or keeps running it in parallel “just in case”
- Who sees which data, and what the agent may write, as against what it may only read
- Whether after the pilot you go to production, fix the data and repeat the pilot, or close the project
- What happens when the agent answers badly: who reports it, to whom, and within what time they get an answer
Five reasons rollouts stall halfway
A stalled rollout rarely looks like a failure. It looks like a project that is “still running”, except nobody remembers where it stopped or whose move it is now. Below are the five reasons we run into most often, how each looks from the outside, and what to do instead.
Four of the five are created before the first line of configuration, while the process is being chosen and the decision rights are being set. That is why this page ends with a checklist rather than a schedule. Adoption is worth measuring from week one, and our own figures with the measurement method and the caveats sit on the AI automation agency page.
| Why it stalls | How it looks from outside | What to do instead |
|---|---|---|
| The process was picked because it hurts most | the project bogs down in exceptions, every case is different | take the most frequent and dullest process, leave the painful one for the second round |
| Data in private spreadsheets and inboxes | the agent answers in generalities, after a week the team stops asking | one source first, with an owner and a stable column structure |
| No threshold at which you say no | the pilot is in its third month, because any result can be defended | write the threshold and the decision date before the pilot, not after it |
| The team hears about it last | the agent works, two people out of fifteen use it | show the team the pilot in week one and collect what they report |
| The scope grows during the pilot | a new system every week, there is never a moment to sign it off | freeze the pilot scope, park the rest on a list for after go-live |
This list comes from our own implementation conversations, not from market research. Adoption can be measured, and it is worth doing from the first week: in our founders' performance marketing agency the agent ran 83 operations in March 2026 with 4 people using it, and close to 63,000 in August with 23 people. Source: the call log in our data warehouse, March to August 2026. That is a usage count, not proof of savings.
What should you expect from an implementation partner?
A good conversation starts with questions about the process, not with a tool demo. You will recognise a bad one because after twenty minutes you already know what you are buying, and nobody has asked how many times a week the task repeats, where the data comes from, or who on your side decides the scope. The same test applies whether you hire an AI agent development company or a freelancer: ask how they will test the agent on your own cases before go-live.
You can ask the six questions below at the first meeting. “It depends on the scope” is a fair answer on the price of a rollout, because the cost follows the number of systems and the number of exceptions in the process, and what settles it is the payback rather than the sticker price, which is the arithmetic in automation ROI. It is not a fair answer on data, on permissions, or on what you get in writing.
- Where do you start: with a workshop and a measurement of the process, or with configuring a tool?
- What data does the pilot run on: real data or sample data?
- Where does our data go, who has access to it, and what stays in the call log?
- What do I get in writing after the workshop if I decide not to go ahead?
- How does pricing work: after you know the process, or as one number at the first meeting?
- How many rollouts are you running at once, and from when do you realistically have time for us?
Common mistakes
Five mistakes made before the rollout even starts
The first process was chosen because it hurts the most
The processes that hurt most are the rare ones, full of exceptions. The pilot bogs down in special cases and the team learns that an AI agent cannot handle the way this company works.
Process time taken from a declaration instead of a measurement
With no “before” number there is nothing to show after 90 days. The conversation about the result ends in an exchange of impressions, and the next rollout does not get a budget.
Data scattered across private spreadsheets and inboxes
The agent answers with whatever it can find, so it answers in generalities. After a week the team goes back to clicking through systems and the pilot dies quietly, with no decision taken.
The process owner named only after the contract is signed
The scope then gets set by somebody who does not do the work. The agent answers questions nobody on the team asks, while the real ones stay unhandled.
The team hears about the agent after go-live
People see a tool that was tried out on them, and they do not use it. A rollout with no users costs exactly the same as one that works.
Checklist
Checklist before you decide to go ahead
Tick what you have today. If more than one line is missing, start with the workshop, not with the integration.
- One process named specifically, repeated at least a few times a week
- The data for that process in a system you can read it out of
- The “before” number: hours, number of repeats, or time from question to answer
- One person with the right to decide the scope of the process, named
- Someone who can grant read-only access to 2 or 3 systems
- The threshold and the date at which you say no and close the project
- The team running the process knows about the pilot and knows where to report bad answers
- The six questions for the implementation partner printed for the first meeting
Sources
Sources
- n8n documentation: the integration list and the ways to pull data out of systems, to check the “data sits in a system” condition
- HubSpot developer documentation: private app scopes, for when you grant read-only access to the CRM
- BigQuery documentation: roles and dataset permissions, for when the process runs on a warehouse
- JustAutomate, AI automation agency: our own two rollouts with the measurement method and the caveats, including the agent usage log
Let us look at your process
Pick the process before you pick the tool
On a free 30-minute consultation we take the five questions from this page and run them against your own example, and we say plainly whether the process is ready for a first pilot or whether the data needs sorting out first. The process workshop costs EUR 3,500 net per week and ends with a process map, metrics and a 30-day plan you can execute without us. The rollout is quoted after the workshop. You write to Paweł, not to a form.
Book a free consultationFrequently asked questions
How do you implement an AI agent in a B2B company?
Start by picking one process: repeatable, with its data in a system, and with an effect you can measure. Then run the five readiness questions and name the person with the right to decide the scope. Only then comes the workshop, a pilot on real data, production, and a measurement after 30 days and after 90 days.
Which process should you pick for the first AI agent rollout?
The most frequent and the dullest one, not the one that hurts most. A task repeated daily, with its data in a single system and a result you can check in a minute. Rare processes full of exceptions belong in the second rollout, once the team has had time to trust the agent.
How do you know a company is not ready for an AI agent yet?
When the data for the process lives only in people's heads and private spreadsheets, when nobody can decide to change the way the work is done, or when the process itself changes every month. In those three cases the workshop still makes sense, because it ends with a map and a plan, but the rollout waits.
Do I need an IT department to implement an AI agent?
No. You need someone who knows the process and can take decisions about it, plus someone who can grant read-only access to 2 or 3 systems. That is usually 2 to 4 hours of work across the whole project. Knowledge of language models is not needed on your side.
Will an AI agent work if we keep our data in spreadsheets?
Yes, as long as the spreadsheet has one owner and a stable column structure. A sheet in Google Workspace or Microsoft 365 is a data source like a CRM or an accounting system. The format is not the problem. Fifteen private copies of the same file in the team's inboxes is.
What if there is no single decision maker in the company?
Name one before you start, even if the role did not formally exist. It has to be someone who knows the process and can say that from Monday it works differently. A steering committee does not replace that role, because questions about exceptions come up daily during a pilot, not once a fortnight.
How do you spot a partner who is promising too much?
By the promise of a rollout in one day, by a ready-made solution for any company, and by a price given before anyone has seen the process. Ask what data the pilot runs on, where your data goes, and what you get in writing after the workshop if you decide not to go ahead.
How much does an AI agent rollout cost?
The only price we publish is the process workshop: EUR 3,500 net per week, usually 1 to 7 weeks. The rollout is quoted after the workshop, because the cost follows the number of systems and the number of exceptions in the process. Model usage is paid to the provider separately, at EUR 0.005 to 0.07 per query.