← All posts Article · Dec 9, 2025

How to Build an Internal AI Agent, or Hire a Team to Build It

Tom van Wees·Dec 9, 2025·16 min read·Updated November 2026
How to Build an Internal AI Agent, or Hire a Team to Build It

Build in-house when you have engineers who will maintain the agent for years and the process touches 1 system you wrote yourself. Hire a team to build AI agents for internal workflows when you need a supervised process running within a few weeks and your people's hours should go on the rules. Either way, the 7 steps below stay the same.

At Lleverage we think the route matters less than who owns the process. A hired team can write the connections, the escalation rules and the monitoring. It cannot decide what your order desk may approve without asking, and it cannot supply the order history the agent learns from. Those stay with you whichever route you take.

We build these agents with manufacturers, wholesalers and logistics companies, usually starting with order management, and our engineers do the build alongside the customer's team. This guide sets out what each route asks of your people, then walks the 7 steps from first process to production. To see it applied to your own documents, book a demo.

Should you hire a team to build AI agents for internal workflows or build in-house?

Build in-house if the process is unique to you, you have engineers to keep it running after the first year, and the agent writes into 1 system. Hire a team if the process is a common shape, such as order intake or supplier invoices, mistakes are costly and you need it live this quarter.

The model is the small part of either route. Most of the work is the layer around it: connectors that keep working when the ERP is updated, retries, an audit trail, permissions, and behaviour when an order arrives as a photo taken at an angle. An in-house team builds and maintains that layer. A hired team brings it.

Build in-houseHire a team to build it
Who writes the codeYour engineersThe vendor's or partner's engineers
What your people provideProcess owner, engineers, ERP access, test dataProcess owner, ERP access, test data, a reviewer
Time to a supervised first processDepends on free engineering capacityTypically within a few weeks
Who keeps connections workingYour engineers, for as long as it runsThe team that built it
Who improves the agent after go-liveYour engineersThe business corrects it, the team updates it
Main riskThe builder leaves and nobody knows the systemPicking a team that has not worked in your ERP
Right forUnique process, strong engineering team, 1 systemCommon process, several systems, little engineering time

The in-house case is stronger than vendors tend to admit. A capable engineering team building against a system it wrote keeps the knowledge in the building. The case weakens in year 2, when the builder has moved to another project and the models underneath have changed twice. We compare the costs line by line in build versus buy for back-office AI agents, and the running costs in what AI agents cost.

A third route sits between the two: a partner builds the first agent on a platform, and your team can learn to build the next ones there. SIG Benelux worked with a partner: SPAIK built its order intake on Lleverage, with Dynamics 365 staying the system of record.

What does your internal team have to provide, and what does a vendor team do instead?

Hiring a team moves the engineering off your desk, while the knowledge of the process stays with your people. These are the questions buyers ask most often when they plan who does what.

Who is the process owner, and why can no vendor replace them?

The process owner is the person who already answers for the outcome: the order desk lead or the accounts payable lead. They decide what the agent may do alone, what it drafts for approval and what it never touches. They also sign off go-live. A vendor can write the rules down, but only the owner can say which rules are right.

What test data does the team need?

A recent set of real inbound documents, a few hundred where volume allows, with what your team entered in the ERP for each one. Add the customer list, the article master and the price lists that applied at the time. The ERP history is the answer key, so the agent can be tested before it touches a live system.

What ERP access does the build need?

A user account for the agent with the roles that match its scope, and an administrator for a few hours to set it up and pull the export. The agent fills the same fields a person fills, through the ERP's own interfaces, so nothing is migrated. Our post on connecting AI agents to legacy and API-less ERPs covers on-premise systems and those without an API.

What does a vendor team do that your team then does not?

It builds the connections to your ERP and mailbox, writes the agent's instructions and escalation rules, runs the test against your history, sets up monitoring and keeps the agent working as your systems change. With us, connecting the ERP and the systems around it is part of the monthly price, including our engineers' work with your team on the first process.

A rollout typically needs 4 people from your side: the process owner, 1 or 2 people who do the work today, an ERP administrator and a sponsor who approves go-live. The process owner carries most of it. We set out the weeks and the hours per role in what an AI back-office rollout looks like.

What is an internal AI agent, and how does it differ from automation?

An internal AI agent reads unstructured input, works out what it means against your own data and business rules, acts inside your systems of record and escalates what it cannot resolve. Conventional automation runs steps somebody specified in advance. The agent works out what the step should be.

That changes what you can automate. A rules engine needs the decision made before it runs, so it handles the orders that arrive clean and leaves the awkward ones to the desk. An agent takes on the reading and the deciding, which is where the hours go.

Conventional automationInternal AI agent
InputFixed format, known field positionsEmail, PDF, scan, spreadsheet, free text
DecisionRules written in advanceWorked out from the document and your history
Unexpected inputFails or routes to a personHandles it, or escalates with its reasoning
EvidenceExecution logThe result plus the source it came from
Who maintains itDevelopers, after each changeThe business, by correcting the agent
Right forHigh volume, zero varianceHigh variance, judgement-heavy intake

The word "agent" now gets used for anything with a language model in it. Our test is narrower: if it cannot act inside your ERP and cannot show why it did what it did, it is a chat assistant. We have written separately on supervised agents versus background automation, the other distinction worth being precise about.

Step 1: Which process should your first internal AI agent handle?

Pick a process that begins with somebody reading something, runs at least 20 times a week, and has a person who already owns the outcome. Order intake, supplier invoices, specification requests and customer questions fit. A process that runs once a month, has no named owner or leaves no record to learn from does not.

The volume floor sets how fast you learn. Below about 20 runs a week, it takes too long to see whether the agent is right, because there is too little traffic to show the error pattern. Above that, a fortnight of running shows where it struggles.

The reading condition rules out the most candidates. If the input already arrives structured from a source you control, a connection or an existing bot will do. The agent earns its place where the input is a mess: hundreds of customers, each with their own purchase order layout, half of them scanned.

History is the condition most often missed. At SIG Benelux, 88% of order lines match the right article first time, including lines with no article number, because years of order history in Dynamics 365 do the translation.

The audit that takes an afternoon

  1. List every process where a person opens an attachment before anything happens. That list is your candidate set.
  2. Count the weekly volume and time 1 run end to end, splitting reading and deciding from keying and filing.
  3. Check what history exists. 2 years of processed records in the ERP is a strong base. A folder of PDFs with nothing structured is a weak one.
  4. Name the person who will review the agent's output. If nobody will, pick a different process.

Step 2: How do you scope what the agent is allowed to decide?

Write down 3 lists before any building starts: what the agent decides on its own, what it drafts for a person to approve, and what it must never touch. Scope sets the agent's authority. Most internal agent projects stall because nobody wrote it down and the gaps showed up in production.

Be specific. "The agent handles order intake" can be read 10 ways. "The agent creates a draft sales order when it identifies the customer, matches every line to an article and finds no price deviation above 2%; otherwise it flags the order with the reason" can be tested.

The escalation rules need the same care as the normal path, because they make the agent safe to run. Decide what happens when the agent is unsure, when 2 documents disagree, when a customer is new and when a value sits outside its usual range. Each case needs a named person, and a generic exception queue that nobody reads does not count.

Put the 3 lists where the reviewer can see them. A page next to the process, in the words the desk uses, works better than a formal specification. Date each version when you widen the scope, so a dispute about what the agent was meant to do has an answer.

Scope also sets the measure. If the agent drafts and a person approves, the measure is how many drafts pass review unchanged. At SIG Benelux, 73% of order emails reach Dynamics 365 as a draft order on their own, and the desk opens, checks and approves it.

Step 3: Who builds it, your engineers or a hired team?

Decide this third, once you know the process and the scope, because both change the answer. A narrow scope in 1 system you wrote favours building in-house. A process that reads from a mailbox, checks against several systems and writes into an ERP you did not write favours hiring a team.

Ask any team you consider to show an agent running in your ERP, or one like it, and to run its first test on your own history rather than sample documents. A test on your export shows the hard cases a demo never reaches.

There is also a split design: hire the reading, keep the writing. The agent turns the incoming document into a clean structured result, and a bot or connection you already own writes it into the system of record. Our integrations cover the common mid-market ERPs directly, but keeping an existing bot on the write step is a sound design.

Step 4: What does an internal AI agent need to work?

4 things: access to the input where it arrives, a source of truth to check against, a write path into the system of record, and a reviewer. The one most often missing is the source of truth, because companies assume the model brings the knowledge. The knowledge is in your own data.

The input is usually a shared mailbox or a folder. An agent reading a personal inbox gets blocked by holidays, permissions and mail rules. A shared address the team already uses for orders avoids all 3.

The source of truth is where most of the quality comes from: customer records, article masters, price agreements, past orders and specification libraries. Oude Reimer brought 170 manuals from more than 15 manufacturers into 1 searchable knowledge base, and troubleshooting questions that meant scrolling through hundreds of pages now get referenced answers in 70 seconds.

The write path decides whether the work is finished or only prepared. An agent that writes a neat summary somebody then retypes has moved the work without removing it. At Topa Bathroom Products, over 90% of incoming orders now go into Business Central with no manual input, and customers get a confirmation within 30 seconds.

The reviewer can be anyone on the desk who knows the work. They see what the agent proposed, approve or correct it, and the corrections feed back. That loop improves the agent and builds the team's trust in it. On Lleverage, someone can also teach the agent a process by showing it on a screen-share call, and skills improve from feedback, with a person approving each change.

Step 5: How do you get an internal AI agent from prototype to production?

Run it alongside the desk on real work before it writes anything. Feed it the live input, let it produce its result and compare that with what the desk did. 2 weeks of this shows the real accuracy and the real exception pattern, which a test set never does.

In our experience teams skip this step most often, and that is why go-live feels like a gamble. Test documents are the ones somebody chose. Real input includes the order that arrives as 3 forwarded emails, the scan with a coffee ring and the customer who puts quantities in the subject line.

SIG Benelux kept its first scope to 1 location, 1 process and the central order mailbox, and ran the automation alongside the team so everyone could see where it failed without any customer noticing.

Once the comparison holds, go live narrow: 1 customer, 1 document type, 1 person reviewing everything. Widen by adding customers before you add autonomy. Xpol put its largest retailer live first, then the supplier set covering roughly 60 to 70% of weekly volume, with long-tail customers added as the rules grew.

"It's a matter of building trust in the organisation with these kinds of initiatives. You can't just throw something like this over the fence." Cees Maaskant, General Manager, Xpol

Xpol's agent prepares a draft that a person signs off, and it leaves ambiguous fields blank instead of guessing. One week into production, the team reported about 20 minutes saved per large order across 150 orders a week, with 25 customer rulesets that had lived in the heads of senior staff.

Step 6: How do you supervise an agent once it is live?

Watch 4 numbers weekly: the share of runs finished without a person, the share escalated, the share corrected after approval, and the time from arrival to finished record. The third matters most. Corrections after approval mean reviewers are waving work through, which is worse than a high escalation rate.

Set the thresholds before go-live, so nobody argues about what good looks like while the agent runs. A workable start: escalations above 20% mean the scope is too wide, and corrections after approval above 2% mean review has become a formality.

The agent scores how sure it is of each piece of work. Below the line you set, a person decides; above it, the work goes through on its own. Only you move that line, and you move it as the numbers earn it.

Supervision should get cheaper over time. The pattern to look for is a process moving from every run reviewed, to spot checks, to exceptions only. At Microtechniek, around 60 purchase orders a week went into Ridder iQ without anyone typing them in August 2026, giving more than 500 hours a year back to the admin desk, and the admin steps in only where the workflow is unsure.

Keep the evidence trail however independent the agent becomes. Every result should trace back to the document and the rule behind it. At SPL Treatments, the agent checks first-time aerospace parts against a library of more than 800 standards, and every recommendation cites the specification section it came from. In a Nadcap-regulated process, that citation lets a person check the route line by line, which is what makes the output usable.

Step 7: How do you go from one agent to a back office that runs itself?

Add the neighbouring process next. If the first agent handles order intake, the next is the order confirmation or the invoice that follows, because it shares the customer data, the reviewer and the trust already built. Each agent that shares context makes the next one quicker to set up.

Most of that gain comes from knowledge. Once article matching works, a quotation agent can use it. Once customer-specific rules are written down for order intake, the delivery agent can use the same ones. At J. Kisch & Zonen, 80% of customer support questions were repeats, and AI now handles the majority of them.

About 3 or 4 agents in, the constraint usually shifts to data quality: article masters with duplicates, customers recorded twice, price agreements kept in a spreadsheet. Fix it then, because by that point the agents have shown which records cause the trouble.

What goes wrong when companies build internal AI agents?

5 failures account for most of it: a process picked for visibility over volume, a scope written as 1 sentence, no history and no reviewer, autonomy granted before breadth, and success measured in keystrokes. Each one shows up whether you build in-house or hire a team.

Starting too big is the most common and the most expensive. A programme that begins with "automate the back office" produces an architecture diagram. A programme that begins with 1 document type from 1 customer produces something live in weeks.

Underestimating the data is second. The model brings the reading, and your records bring the answers. A messy article master shows up on day 1, which helps, though nobody budgeted for it.

Going live on autonomy instead of breadth damages trust fastest. An agent with full authority over a narrow slice will meet an input nobody expected, and with no person in the loop, the first anyone hears is a wrong order at the customer. The same agent running supervised across a wider slice meets that input too, and a person catches it.

Measuring keystrokes flatters any automation. The number that decides whether it was worth it is hours returned to named people and what they did with them. At Topa, nearly 4 FTE (3.8) moved from order entry to after-sales support and service planning.

Frequently asked questions

When is building an internal AI agent in-house the better choice?

Build in-house when the process is unique to you, the agent writes into 1 system your engineers know, and they will maintain it for years. Hire a team when the process is common, such as order intake or invoices, and it touches several systems. Either way, your process owner, test data and ERP access decide whether it works.

How long does it take to build an internal AI agent?

With a hired team on an existing platform, the first process typically runs supervised within a few weeks. Most of that time goes on scoping, data access and running alongside the desk. A custom in-house build depends on the engineering time you can free, and maintenance starts the day it goes live.

What does our team need to provide if we hire a team to build the agent?

A process owner who decides the rules, 1 or 2 people who do the work today, an ERP administrator for the account and an export, and a sponsor who approves go-live. Add a few months of real documents with what your team entered for each. The hired team handles connections, testing and monitoring.

Can an internal AI agent work with an old ERP?

Usually, including systems with no API, though the write path takes more work. SPL Treatments runs a legacy desktop ERP with no API: the agent drafts the route with every recommendation cited, and a person keys the confirmed route in by hand. Where an ERP allows it, the agent fills the same fields a person fills.

Should the first internal AI agent be fully autonomous?

No. Start with a named person reviewing every run, and widen the input before you widen the authority. Autonomy should follow a measured error rate. A wrong order that reaches a customer in the first week costs more trust than a month of reviewed drafts, so let the numbers move the confidence line.

Start with your own documents

Whichever route you choose, the decisive work comes before any building: picking a process that begins with reading, and writing down where the agent's authority ends. If you hire a team, judge it by how it tests against your history and who keeps the agent working in year 2.

Send us 3 weeks of real orders, invoices or specification requests as they arrive. We will show which share an agent resolves on its own, which share it escalates, what your team would need to provide and what the first scope should be. Book a demo and we will work through your documents instead of a sample set.

Give your back office an AI workforce.