An AI back-office rollout of 8 weeks takes one process from kick-off to go-live. Weeks 1 and 2 map the work and test the agent on an export of past documents. Weeks 3 and 4 connect the ERP. Weeks 5 and 6 run supervised beside your team, and weeks 7 and 8 let routine cases run on their own.
Two questions come up in almost every first conversation we have with an operations director. They want to know how long it takes until the agent is live, and how many hours their own team has to give. This guide answers both, week by week, for one process in one ERP. At Lleverage we think the first proof should run on your own data before anyone opens a connection to the ERP, because a test on last quarter's orders settles more than any demo. We set no fixed number of weeks for a customer. The first process typically runs supervised within a few weeks, built by our forward deployed engineers with your team. 8 weeks is a fair frame for planning one process up to the point where routine work runs without a person.
The examples below come mostly from order intake, the process most companies start with. The same shape holds for invoice matching. To see what the agent does once it is live, look at order management or accounts payable. Or book a demo and bring a week of your own documents.
How long does an AI back-office rollout take?
For one well-chosen process in a mainstream ERP, plan on 8 weeks from kick-off to go-live. The agent works on real documents under supervision from around week 5. Most of the time goes into mapping, connecting and supervised running, the work that makes the result safe to rely on.
The 8 weeks split into 4 blocks of 2. Each block ends with a decision your team takes, based on what the agent did with your documents.
| Weeks | What happens | What your team decides at the end |
|---|---|---|
| 1 and 2 | Kick-off, interviews, process map, test on an export of past documents | Whether the results on your own history justify connecting the ERP |
| 3 and 4 | ERP access, rules per customer or supplier, test writes in a safe environment | Whether the agent fills the fields the way your desk does |
| 5 and 6 | Supervised running on live documents, every result checked by a person | Where to set the confidence line for the first routine cases |
| 7 and 8 | Routine cases run on their own, exceptions go to a named person | Whether the process is live and who owns it from here |
Two things move that schedule more than anything else. The first is access to the ERP: an on-premise system with no interface for outside software can add weeks before any build starts. The second is scope. One mailbox and one document type fit in 8 weeks. Three mailboxes, two countries and a phone line do not.
Some projects finish faster. Our partner SPAIK ran SIG Benelux's order intake project on Lleverage as a 4-week Kickstart, scoped to one location, one process and the central order mailbox. That is one customer's figure on a deliberately narrow scope, and it shows what scope discipline buys.
Who from your team does the rollout need, and for how many hours?
A rollout needs 4 people from your side. They are the process owner, 1 or 2 people who do the work today, someone with admin rights in the ERP and a sponsor who approves the go-live. The process owner carries most of the load, at around 2 to 4 hours a week.
The hours below are a planning budget, the time we suggest you reserve in diaries before kick-off. The real figure depends on volume and on how much of the process already exists in writing.
| Role | Weeks 1 and 2 | Weeks 3 and 4 | Weeks 5 and 6 | Weeks 7 and 8 |
|---|---|---|---|---|
| Process owner (order desk lead, AP lead) | 4 to 6 hours: kick-off, interviews, review of test results | 2 to 3 hours a week: rules and edge cases | 2 hours a week: review meetings | 2 hours a week: go-live decision, handover |
| People who do the work today | 1 to 2 hours each: interview, show how a hard case is handled | 1 hour: check a sample of results | Inside the normal working day: checking drafts instead of typing | Inside the normal working day: handling exceptions |
| ERP or IT administrator | 2 to 4 hours: the data export | Half a day: user account, roles, test environment | 1 hour: monitoring | Little or none |
| Sponsor (operations or finance director) | 1 hour: kick-off and scope | None | 30 minutes: progress review | 1 hour: go-live decision |
During supervised running, the people who did the work check the agent's drafts in the time they used to spend typing them. That review time replaces work they already did, so it adds little to the week.
The hours that matter most are the ones in weeks 1 and 2. The rules that make a process work usually live in the heads of 2 or 3 long-serving people. An hour with each of them early saves days of corrections later. If those people cannot free an hour in the first fortnight, the rollout will slip.
What you do not need is a project team. There is no data migration and no change to the ERP. The agent fills the same fields a person fills, through the ERP's own interfaces, so the system checks its work the same way it checks a colleague's. More on how that connection works is on our integrations page.
How do you run a proof of concept on a data export before any live integration?
Export 3 months of inbound documents with what your team entered in the ERP for each one. Let the agent process the documents, then compare its output field by field with your team's. The ERP history is the answer key, and nothing touches a live system.
We think buyers should insist on this step, with us or anyone else. A test on your own export shows the hard cases a demo on sample documents never reaches. Think of the customer who sends orders as a photograph, or the supplier who changed their invoice layout in August, or the article descriptions only your desk understands.
What goes into the export
For order intake, the export holds 4 things:
- The inbound documents: order emails with their attachments, for a recent period of about 3 months, ideally a few hundred of them.
- The result for each one: the sales orders your team created from those emails, with customer, lines, quantities, prices and delivery dates.
- Master data: the customer list with delivery addresses, the article list and the price lists that applied in that period.
- A short list of the rules nobody wrote down, collected in the interviews: which customer always omits the delivery week, which one orders in boxes while the ERP counts pieces.
For invoice matching, the same pattern holds: supplier invoices, the purchase orders and goods receipts they belong to, and how each invoice was finally booked.
An administrator can usually pull the ERP side with the standard export or reporting tools in a few hours. The mailbox side is often the harder part, because attachments have to come out with the emails they belong to. Plan for it in week 1.
What the test measures
The agent processes every document in the export. Then the two results sit side by side: what the agent would have created and what your team created. Three measures come out of it.
- Document level: the share of documents where every field matches what your team entered.
- Line level: the share of order lines matched to the correct article, the measure that decides most of the review work later.
- The failures, by cause: which customers, layouts or rules account for the documents that did not match.
The failures by cause decide the plan for weeks 3 and 4. A rate of 70% with failures spread evenly across customers needs a different plan from 70% where 3 customers account for nearly all the misses. The first tells you something about the data. The second gives you a short list of rules to add in weeks 3 and 4.
A disagreement is sometimes the agent being right. Exports regularly turn up orders your team keyed with a typo or an old price. Those cases are worth a look, because they show what the review step during supervised running will catch.
Why the history usually holds the answer
The question we hear most before a test is whether the data is clean enough. At SIG Benelux, roughly 3% of orders carried an article number and about 47% had only a description. Years of order history in Dynamics 365 already contained the translations the sales desk made by hand, and 88% of order lines now match the right article first time.
"I knew the technology made this possible. My biggest question was the quality of our data. It was a positive surprise to see the order history was enough of a base to get results this fast." Ditte Smit, Marketing Director, SIG Benelux
The export also answers a privacy question early. Order emails carry names and contact details, so treat the export as personal data and agree where it is processed before it leaves the building. Lleverage hosts in the EU by default and works under GDPR, and the access rules are set per connection, as described on our control page.
What happens in weeks 1 and 2?
Weeks 1 and 2 turn a process that lives in people's heads into a written map, and test the agent on the export. By the end of week 2 you hold 2 things: a process map with every exception your team named, and a result on your own history that tells you whether to connect the ERP.
Kick-off, day 1. One meeting of about an hour with the sponsor, the process owner and our engineers. It fixes the scope in one sentence, for example "orders arriving in the central order mailbox, for the Netherlands, created as sales orders in Business Central". It also names the success measure and the person who will own the process after go-live.
Interviews, days 2 to 5. Our engineers sit with the people who do the work and watch them handle real documents. The useful questions are about the hard cases: what makes an order take 15 minutes instead of 2, which customers need a phone call, what happens when an article code is unknown. SIG Benelux's first 2 weeks went into 4 in-depth interviews with the sales desk, a strategy session with the management team and reading along in the real order mailbox.
The export, week 1. The ERP administrator pulls the history and the master data described above. If the export is ready by the end of week 1, the test runs in week 2.
The test, week 2. The agent runs over the export, the results are compared and the failures are sorted by cause. The process owner reviews them in a session of an hour or two.
At the end of week 2 the team decides whether to connect the ERP. If the export shows that the documents are readable and the failures cluster around rules you can name, connecting the ERP is the next step. If the failures are spread across everything, the scope is probably too wide, and the right move is to narrow it to the customers or suppliers that cover most of the volume.
What happens in weeks 3 and 4?
Weeks 3 and 4 connect the agent to the ERP and teach it the rules the export exposed. The agent gets its own user account with the rights your desk has for this process. Test writes run in a safe environment until the fields come out the way your team would fill them.
Access. The ERP administrator creates a user for the agent, with the roles this process needs and nothing more. That account decides what the agent can see and change. For SAP, our team sets the connection up with your SAP team, as described on our SAP integration page. An on-premise ERP is reached without opening it to the internet. Where an ERP has no interface at all, the agent works the screen with the access a person has.
A test environment. Many ERPs offer one. Microsoft includes 1 production environment and 3 sandbox environments in the Business Central Premium and Essential subscriptions at no extra charge. A sandbox can be created as a copy of production. In that copy the job queue tasks are stopped and the integration settings cleared, so a test write does not set off a live process. Test sales orders created there show the process owner exactly what would land in production.
Rules per customer or supplier. The failures from week 2 become rules. Some are conversions (this customer orders in boxes of 12). Some are routing (orders above a certain value go to a named person). Some are knowledge (this description from this customer means that article). Xpol, a fresh flower wholesaler, ended up with 25 customer rulesets, including unit conversions, weekday-specific label text and multi-depot splits that had lived in the heads of senior staff.
A sample check. At the end of week 4, the people who do the work check a sample of the test orders. Their verdict decides whether supervised running starts in week 5.
These are also the weeks where the ERP side can stretch the plan. At Microtechniek, the on-premise Ridder iQ installation lacked the API module the cloud version has, so access had to be solved before anything else. A Ridder specialist installed the module on site in December 2025, and the order work was built on top of it. If your ERP is old or heavily customised, ask about this in week 1, so it does not surface in week 4. Our guide to connecting an AI agent to your ERP goes deeper on the options.
What happens in weeks 5 and 6?
In weeks 5 and 6 the agent works on live documents and a person checks every result before it counts. The agent prepares the sales order or the invoice match, your team reviews it, approves it or corrects it, and every correction becomes part of how the agent handles the next similar case.
This stage is slow on purpose: it produces the evidence for the go-live decision, and the corrections that make the agent better. A process owner who has seen several hundred results, with the reasoning behind each one, can decide where the work may run on its own.
There are 2 ways to run these weeks, and the choice depends on how much risk the process carries.
| Approach | How it works | When it fits |
|---|---|---|
| Shadow running | The agent processes documents alongside the existing process, the team keeps working as before and the results are compared | High-risk processes, or when the team is wary and needs to see results without changing anything |
| Supervised drafts | The agent creates drafts in the ERP and a person checks and approves each one instead of typing it | Most order intake and invoice matching, once the export test and the sample check went well |
SIG Benelux used the first. Its second fortnight combined building with shadow running, so the team saw every miss while customers saw none. Xpol made review a permanent part of the design. The agent prepares a draft that a person signs off with one click, and a field it cannot resolve stays empty for that person to fill.
During these weeks, watch 3 numbers:
- The share of results approved without a change. It should climb week on week as rules are added. A flat line means corrections are not being captured.
- The exception mix. Which cases still need a person, and whether that list is shrinking or only changing shape.
- Review time per document. How long a check takes compared with typing the document from scratch.
Hold one review meeting a week, 30 to 60 minutes, with the process owner and our engineers. Go through the corrections, decide which become rules, and look at the cases where the agent was unsure. Our piece on supervised agents and background automation explains when each mode is the right one.
What happens in weeks 7 and 8?
In weeks 7 and 8 the routine cases start to run on their own. The exceptions go to a named person with the context attached. Go-live is a decision your team takes on the evidence from weeks 5 and 6. It covers the cases the agent has shown it handles, and everything else stays with a person.
Setting the confidence line. The agent scores how sure it is of each piece of work. Below the line, a person decides. Above it, the work goes through on its own. Your team sets the line, starting cautiously, and only your team moves it. A common first step is to let the customers or suppliers with the best record in weeks 5 and 6 run on their own while the rest stay supervised.
Phasing by volume. Going live does not have to mean every customer on the same day. Xpol went live in production on 20 April 2026 with its largest retailer first, then the supplier set covering roughly 60 to 70% of weekly volume over the following weeks, while long-tail customers were added as the rule library grew.
Handover. By the end of week 8, 5 things should be in place:
- A named owner for the process, usually the person who ran it before, with the right to move the confidence line.
- A named person for exceptions, with a turnaround the desk agrees to.
- An audit trail your auditor can read: what the agent did, on which document, and why.
- A weekly view of what the agent handled and what it held back.
- A route for changes, so a new customer layout or a new rule reaches the agent the same week.
Early wins to expect. The first visible change is often outside office hours. At Microtechniek, orders were entered only when one person was at their desk. Order entry now runs in the evenings and at weekends, and orders sent late on a Friday are waiting in Ridder iQ on Monday morning. At Topa Bathroom Products, more than 90% of incoming orders are processed into Business Central with no manual input, and the team came back from the year-end break to find the backlog cleared by the end of Monday.
Go-live is also where the work of the next process starts to shrink. The ERP connection, the access model and the rules your team taught the first agent are already in place for the second, which is why the second process is faster. Our article on moving from a dozen AI pilots to one autonomous back office covers that step.
What makes a rollout slip past 8 weeks?
Rollouts slip for 5 reasons: the scope grows, the ERP has no way in, the export arrives late, the experts have no time, or nobody owns the exceptions. Each one can be spotted in week 1.
| Cause of delay | Early sign | What to do in week 1 |
|---|---|---|
| Scope grows | A second mailbox, country or document type appears in the kick-off | Write the scope in one sentence and park the rest for process 2 |
| No way into the ERP | On-premise system, heavy customisation, no interface for outside software | Ask the ERP partner what interfaces exist and book them early |
| Late export | Nobody owns the mailbox or the ERP reports | Name the person who pulls the export, with a date |
| No time from the experts | The 2 people who know the rules are fully booked | Block their interview hours before kick-off |
| No exception owner | The go-live plan says exceptions go to "the team" | Name one person and a turnaround |
A rollout scoped to one mailbox becomes a rollout scoped to every channel the moment someone in the kick-off mentions the phone orders. Finish one document type first. Everything else becomes process 2, which starts faster because the connection already exists.
An agent that handles most of a process creates a new queue holding the rest. If nobody owns that queue, exceptions pile up, the desk loses trust and the go-live gets reversed. Name the person before week 5.
The cost question belongs in week 1 as well. Lleverage charges one monthly price per agent, with integration, AI usage and ongoing improvement included, and we price after we have seen the work. Our pricing page explains how that is scoped, and our article on what an AI agent costs sets out the wider picture.
Frequently asked questions
How many hours does an AI back-office rollout take from our team?
Plan for about 2 to 4 hours a week from the process owner across the 8 weeks. The people who do the work give 1 to 2 hours each in the first fortnight. The ERP administrator gives half a day to a day for the export and access. During supervised running, checking drafts replaces typing, so it adds little to the desk's week.
Can we test an AI agent without connecting it to our ERP?
Yes. Export about 3 months of inbound documents with what your team entered in the ERP for each one, plus the customer, article and price lists. The agent processes the documents and its output is compared field by field with your team's. Nothing touches a live system, and you decide on the result whether to connect.
What does go-live mean for an AI agent?
Go-live means the routine cases run on their own and the exceptions go to a named person. Your team sets the confidence line: below it a person decides, above it the work goes through. Many teams go live per customer or supplier group, starting with the ones that cover most of the volume and adding the rest later.
Do we need IT for an AI back-office rollout?
Only for a few hours. The ERP administrator pulls the data export, creates a user account for the agent with the right roles, and sets up a test environment where the ERP offers one. Nothing is migrated and the ERP itself is left unchanged. The process owner in operations or finance runs the project.
Which process should go first in an 8-week rollout?
Choose a process with high volume, documents in many formats and a result that lands in one system, such as order intake or invoice matching. Volume gives the agent enough cases to learn from. A single destination makes success simple to measure. Avoid a process that touches 6 systems or happens once a month.
Plan the first 8 weeks with us
Before any kick-off, do 3 things: pick one process and write its scope in one sentence, name the person who will own it after go-live, and ask your ERP administrator whether they can export 3 months of documents with the matching ERP records. With those in hand, week 1 starts on the first day.
Our forward deployed engineers then sit down with your team and map where the manual work and the exceptions sit, and we think the first test belongs on your own history. Book a demo and bring a week of orders or invoices, including the awkward ones. We will go through them with you and sketch what an 8-week plan looks like for your process.