Skip to content

GuidesAI at work

BoondManager and AI: what IT services firms can delegate, task by task

BoondManager and AI: Boond’s MCP connector, API and an AI employee. What an IT services firm delegates on timesheets, bench and skills files, and what it keeps.

20 min read

In BoondManager, AI can already take on a good share of an IT services firm’s repetitive work: flagging missing timesheets before invoicing, matching assignment end dates against open requirements, preparing a skills file on the client’s template. It gets there through three doors: the AI Boond builds in itself, the MCP connector that plugs Claude or ChatGPT into your data, and an AI employee that works through the API. Whichever door it comes through, it should never approve a timesheet, put a consultant forward or send anything to a client without a person having decided it.

Boond, the name BoondManager now goes by, claims more than 2,000 IT services and consulting firms as customers. For many of them it is the tool they open in the morning and close at night, where the requirement, the positioning, the assignment, the timesheet and the invoice all live. The question is no longer whether to put AI on it, since Boond ships some itself, but through which door, for which tasks and with whose permissions.

This guide draws on Boond’s help centre, as read on 13 November 2026, and on what we do with Balt. We should say so up front: Balt is a Boond partner and is listed in its partner directory, and we have tried to be as precise about Boond’s tools as about our own.

What does BoondManager handle for an IT services firm?

BoondManager is an ERP built for IT services and consulting firms. It brings sales, with accounts, contacts and client requirements, and recruitment, with an applicant tracking system, its candidates and their positionings, into a single database. Alongside them sit resources (consultant records, assignments, end-of-assignment alerts), activity (time, expenses and absences, what French firms call the CRA) and invoicing, e-invoicing included. Boond sells four plans: Core for up to five managers, Advanced from five to fifty, Enterprise beyond that, and Business, which keeps only the CRM and applicant tracking for a company that already runs another ERP.

That concentration makes Boond ideal ground for an agent, and for the same reason ground where stale data is expensive. An account manager who knows their clients is wary of an end date never corrected or a lost requirement still marked open. An agent reasons on top of it and returns a result that is credible and wrong, which is why you clean data at the moment the agent touches it rather than the whole database before you start.

Which doors does AI use to reach Boond?

Three doors lead AI to your Boond data, and they do not serve the same need. The first is Boond’s own AI, whose product page highlights AI matching between skills, availability and client requirements, along with automatic generation of skills files from your templates.

Then there is BoondPulse, an assistant in beta on the Advanced, Business and Enterprise plans, opened from the orange star at the top of the screen or from a browser extension. It suggests actions based on the screen you are on, such as “Late timesheets”, “Who is available?” or “Follow up on unpaid invoices”. Boond sums up its rule in one sentence: BoondPulse suggests, you validate, and no action is taken on your behalf.

The second door is the MCP connector, published by Boond, which plugs your own assistant into your data. The third goes through the API: a vendor builds an app, your administrator installs it from Boond’s Marketplace, and that is how an AI employee such as Balt works.

Built-in assistant (BoondPulse) Your assistant through MCP An AI employee (Balt)
Where you talk to it In Boond or in the extension In each person’s own assistant In Teams, Slack, WhatsApp, by email or by phone
Who it works for The person who opens it The signed-in user, with their permissions The whole team, with the permissions you grant it
What sets it off A request or a quick action A request, or a scheduled task if the assistant allows A request or a routine
Writing to Boond After you validate Approval by default, which the administrator can disable Your approval is asked before every write, by default
A message to outsiders Drafted, ready to send Sent after your explicit confirmation Never without approval, and no setting removes it

The table places each door rather than ranking them. A built-in assistant serves the person in front of their screen, and an assistant plugged in through MCP serves someone who wants to query Boond from the tool they already work in. An AI employee serves a team handing over recurring tasks that straddle Boond and its other tools.

What does Boond’s MCP connector allow, and with what permissions?

Boond’s MCP connector lets Claude, ChatGPT or any compatible assistant search, read, create and update records, approve timesheets and send emails, with the signed-in user’s permissions. It is included in the Advanced, Business and Enterprise plans, but not in Core. An administrator installs the “MCP Boond” app from Administration > Apps > Marketplace and sets its visibility, then each user adds a connector pointing to https://ui.boondmanager.com/api/mcp in their assistant and authorises it.

Boond announces twelve tools and its help centre describes ten of them, six for reading and four for acting. On top of those come business guides the assistant loads on demand, such as the one for logging an action on the right record or the one for creating a delivery on a project.

Family Tools What they allow
Read search, fetch, describe, dictionary, perimeter, context Search every module, from candidates to invoices, read a record tab by tab, understand your configuration and your custom fields
Act create, update, transition, send_email Create or update a record, approve or reject a timesheet, an absence or an expense report, send an email logged on the record
Missing no deletion tool Nothing gets erased, and invoices, orders and purchases can be updated but not created

Permissions are the best-designed part of the connector. The assistant only sees and changes what the user could see and change themselves, those permissions are checked on every call, and Boond does not keep the conversation. Writes go through manual approval by default, which the administrator can turn off, and Boond recommends leaving reads open and requiring approval for anything that writes.

Two limits are written into the documentation: the connector does not yet do bulk processing or aggregated reporting across the whole database, and it does not read documents attached to records, CVs included. A third point, less visible, matters more if you want to automate. For scheduled tasks or autonomous sessions, Boond warns that the necessary permissions must be switched on, otherwise the tasks stay blocked waiting for someone to click “authorise”. A task that runs overnight therefore pushes you to open up the writes you had put behind approval, at the very moment nobody is watching.

What should you check before plugging AI into Boond?

Four questions apply to every door, Boond’s own connector included, and they belong before the connection rather than after the first incident.

Who supplies the server, and what can it reach?

An MCP server should be treated as a supplier: it holds access to your systems, and it feeds the model text that the model reads as instructions, from tool descriptions to business guides. For the official connector, that supplier is Boond, which already holds your data, and that is the best case there is. Its guides are still text that changes on Boond’s schedule without you having changed anything, and our article on MCP servers sets out the three questions to ask before connecting one.

Community servers for Boond exist too. One of them, open source, exposes 238 tools, deletion included, and depending on the mode accepts a per-user OAuth token, a user token combined with the administration’s client token and client key, or a username and password. Its documentation is careful, but it adds a supplier to the chain. Boond itself gives a practical argument for a declared app. Calls from a partner app connected through OAuth2 or JWTApp do not count against your API quota, whereas calls from a tool that uses a password or the client key do.

Whose permissions does the AI act with?

AI should act with the permissions of the person asking it for something, never with an account of its own that is broader than anyone’s. The test fits in one question: do two people with different permissions get the same answer to the same request? If they do, the tool sees more than one of them is allowed to, and an identity of the agent’s own becomes a path around your access controls.

The MCP connector passes that test by design, since everyone connects with their own Boond account. Boond says as much in its API best practices: a service account suits a background integration with a defined scope, but for an app that relies on the signed-in user’s permissions it would bypass Boond’s rights configuration. That leaves scheduled runs, which have no requester, and which we tie to the permissions of the person who set them up, even if that means they stop when that person leaves.

What does Boond let a third-party tool do?

What a piece of software lets a third party do decides what AI can do in it, far more than the AI it builds in. Boond has all five properties we ask of an applicant tracking system or CRM before connecting an agent. Its API exposes the business objects for writing, its activity log lists creations, updates and deletions, permissions are set per person, history can be extracted and custom fields can be reached.

Its documentation flags three points to plan for. Webhooks are delivered “best-effort”, with no retry, so a missed event remains a silent gap until a daily pass over the activity log catches it. Time entries, absences and activity reports cannot be filtered by modification date. And Boond announced rate limiting from the end of September 2026, on top of a monthly quota that depends on the plan.

Where does the data go?

Boond’s data is full of names: candidates, consultants and client contacts. Our rule fits in one line: what contains no candidate and no client can go through a consumer assistant, and everything else goes through a tool the company has chosen, put under contract and entered in its record of processing activities. The connector’s FAQ says the same thing in its own way, since what the assistant does with the data after reading it, training included, depends on your contract with its vendor.

A connector plugged into an account manager’s personal account therefore takes client records outside the company’s perimeter, and that assistant’s memory belongs to that person’s account. The limits of ChatGPT for recruiters spell out what this implies, including why removing a name is not enough to anonymise a profile.

What does an AI employee add to the MCP connector?

What an AI employee adds is less about access to data than about how the work around it is organised. The connector and the employee read the same database, and either can draft a follow-up or summarise an account. The difference comes down to four points, which start to matter once a task comes back every week and involves several people.

The work starts without anyone asking. Every Monday morning, Balt matches the open requirements in Boond against the consultants rolling off assignment, and suggests the placements you could make in the team’s channel. A routine reports even when it found nothing, otherwise nothing tells it apart from a routine that never ran.

The context belongs to the team. With the connector, everyone has their own assistant, their own conversation and sometimes their own memory. Balt is a single colleague for the whole company: the account manager and the recruiter see the same answer in the same channel, and what he has learned about how your firm keeps its records stays with the firm when someone leaves. Facts about people stay in Boond, which is built to hold them.

Approval lives in the product. By default, every write Balt makes to Boond waits for your approval, and nothing he prepares for a client or a candidate leaves until a person has released it, with no setting to lift that rule. The connector’s approval is a good default, but it can be switched off, and Boond’s documentation itself says when people end up doing so.

Calculations cover the whole database. Balt works by writing code, so a margin per client over the quarter or the list of every late timesheet is a calculation run across all the records, with the source of every figure. That is precisely what the connector does not yet cover.

The connector still holds real ground. To query your own scope from the assistant you already work in, summarise an account before a meeting or find the Java candidates available around Lyon, it does what you ask, at no extra cost on the Boond side. An AI employee costs credits, from €99 a month, and earns its place when tasks recur, cross several tools and serve a team rather than one person. The two sit side by side without trouble, since in both cases Boond remains the system where the facts live.

To connect it, an administrator installs Balt from Boond’s Marketplace and sets its visibility, then the team writes to him in Teams, Slack or WhatsApp. The BoondManager integration page describes what he does there and the permissions he asks for.

What can you hand over to AI on Boond, task by task?

The table sums up the split we recommend for six tasks that come back every week in an IT services firm. Each is detailed below with what the AI does badly, because that is what decides where the human stands.

Task What the AI does What you keep
Timesheets List the missing ones, spot inconsistencies, draft the reminders Approval and the relationship with the client
Skills file Build the file on the client’s template from the consultant record The consultant’s review and the send
Bench Match end dates against open requirements, explain each match The conversation with the consultant and the decision
Invoice follow-ups Read each file, set disputes aside, draft the follow-up The tone, the dispute and the formal notice
Tenders Gather references, profiles and documents, prepare the decision grid The decision to bid, the price and the signature
Sales pipeline Log activities, flag requirements that have gone quiet Anything that closes an opportunity or drops a candidate

Timesheets, before invoicing

The timesheet is a good first task, because it comes back every month on a fixed date and a result can be checked in a minute. In the last working days of the month, the AI lists the consultants whose timesheet is not entered or not approved, counts the timesheets received against the assignments under way and drafts the reminders, which the manager releases in one pass. It then checks what comes in: days entered outside the assignment dates, a total that does not match the purchase order, an absence booked on a day reported as worked.

It goes wrong on anything that is not written down in Boond. If nobody noted that the client closed its site between Christmas and New Year, it will flag as an anomaly what was a verbal agreement. And since time entries cannot be filtered by modification date, a serious check rereads the current and previous month, then consults the activity log to catch an older correction.

Approval stays with the manager, even though the MCP connector can do it, because an approved timesheet triggers the invoice and the client often signs it on its side. What the AI brings here is more modest and safer: invoicing no longer goes out late for want of a timesheet chased in time.

The skills file, from the consultant record

Boond already generates consistent skills files from your templates, and that is a solid base. What AI adds is fitting the file to one specific requirement. It starts from the technical data on the consultant record and from the client’s requirement, picks the experience that matters for this assignment, orders it and rewords it in the client’s vocabulary, on the client’s template, anonymising the file when needed. It can also return the list of what the profile lacks against the requirement, which no document template does.

The risk is serious, because AI that writes tends to round up, and a technology brushed against on one project quickly becomes expertise. With a career history you select, order and rephrase, and you never add, and the line can be tested with one question: can the consultant defend every line at the client’s first technical question? Our article on the skills file as a sales document sets out that rule, and Cobalt, our sister company, lists the skills file mistakes that lose assignments.

The file leaves the company, so the consultant reviews it before a person releases it. Technical data that is two years old produces a skills file that is wrong and well written, which is more dangerous than a badly written one because no one thinks to doubt it.

Bench time and assignment end dates

Every Monday morning, the AI matches open requirements against the consultants on the bench and those rolling off in the coming weeks. Each match carries its own sentence: what fits the requirement, what is missing, what it could not verify. A sentence can be challenged whereas a score gets applied as it stands, which is why we prefer the first for anything that concerns a person.

Its weak point is the date. The one your ERP holds is almost always the contractual date, and an AI that takes it at face value discovers the roll-off on the day it happens. The delay that costs most is the discovery delay, ahead even of the redeployment delay, as our analysis of bench time shows, and it only shrinks when account managers confirm the dates.

You keep the conversation with the consultant, who may no longer want that technology or a two-hour commute, and the decision to put them forward. The same reading of end dates serves renewals, where margin is lost because nobody noted the date: the AI keeps the calendar and the margin per consultant, and the discussion with the buyer stays human.

Invoice follow-ups

For a follow-up, the AI first reads the invoice in Boond, its due date, the payments received and the account’s history, then sets disputed files aside and matches the tone to how late the payment is. In IT services, a late payment often has an administrative cause that this reading reveals, such as a purchase order number missing from the invoice or a timesheet the client never signed, and a follow-up that ignores it damages the relationship for nothing.

It knows nothing of what was not logged: a transfer promised over the phone for Friday does not exist for it unless it is noted on the record. The tone with a strategic account, the dispute and the formal notice stay with you, and each follow-up goes out after your approval, one file at a time, following our guide to chasing an unpaid invoice.

Tenders and proposals

A tender response takes 25 hours of writing on average according to Loopio’s 2025 survey, and a good share of those hours goes into gathering what Boond already holds. The AI prepares the references drawn from past projects, the consultants available at the start date with their skills file on the requested template and the administrative documents, then pre-fills the decision grid with facts.

The trap is the order, because a grid filled in after deciding to bid justifies the decision instead of making it, and AI makes that habit very comfortable. Nor does it know whether the consultant being proposed agrees, and a profile only belongs in the response if the consultant has been told, has accepted and will be available at the start.

The decision to bid, the price and the signature stay human. The daily rate is worked out from cost and defended with a floor written down before the meeting, as our method for setting a consultant’s daily rate explains. A tender is often won before it is written, which our article on IT services tenders develops.

A sales pipeline that stays current

A pipeline decays through small omissions: the call never logged, the lost requirement still open, the positioning whose status was never updated. After a call, the AI files what was said on the right record as an activity and suggests the update to the requirement or the positioning, which you accept in one gesture. Every week, it flags the requirements with no recent activity and the accounts that have stopped ordering.

It does not decide what a status means. Marking an opportunity as lost, closing a requirement or pulling a candidate from a positioning closes a door, and a decision that closes a door stays human even when it never leaves the company. Duplicates call for the same caution, because a merge cannot be undone, and Boond’s connector has no deletion tool in any case.

What should AI never do on Boond?

AI should do nothing on Boond that leaves the company without approval, and nothing that closes a door without a human decision. The first rule covers skills files sent out, follow-ups, proposals and messages to candidates; the second covers approving a timesheet, rejecting a candidate, closing a requirement or setting a price.

The objection is fair, since a rushed account manager also closes requirements without checking and approves timesheets without reading them. Yet their mistake does not repeat itself identically across three hundred records, and you can always ask them why, whereas a badly set rule in a tool applies the same mistake everywhere, silently. A limit holds when it is built into the product rather than written into an instruction, like the absence of any deletion tool in Boond’s connector, or the absence of any setting that would let Balt write to a client without your approval.

Where should you start with AI on Boond?

Start with what can be checked at a glance and writes nothing. For an IT services firm of thirty consultants, that means two uses during the first month: the check on missing timesheets in the last days of the month, and assignment end dates matched against open requirements on Monday morning. A mistake shows immediately, it costs nobody anything, and the team learns to work with the tool during the three to four weeks it takes for delegation to become a habit.

Then come internal, reversible writes, such as the activity logged after a call, and after them skills files and follow-ups, which leave the company and therefore go through approval. The rule of order is the one in our article on delegating to an agent: you delegate first what can be checked, not what takes the most time.

That is the order we follow with Balt. He works in Boond and in the rest of your tools, Outlook, Word or Turnover-IT, reads what he needs, asks before writing and lets nothing go to a client without a person having decided it. He starts at €99 a month, with 20,000 free credits when you sign up, and our page for IT services companies describes what he takes on there.

Frequently asked questions

  • Does Boond have its own AI?

    Yes. Boond highlights AI matching between candidates and requirements, automatic generation of skills files from your templates, and BoondPulse, a built-in assistant in beta on the Advanced, Business and Enterprise plans. BoondPulse suggests actions based on the screen you are on and takes none on your behalf.

  • How do you connect ChatGPT or Claude to BoondManager?

    Through Boond’s MCP connector. An administrator installs the “MCP Boond” app from Administration > Apps > Marketplace and sets its visibility, then each user adds a custom connector pointing to https://ui.boondmanager.com/api/mcp in their assistant and authorises it. You need a Boond Advanced, Business or Enterprise plan, and an assistant plan that allows custom connectors.

  • Does an AI connected to Boond see more than the user does?

    Not with Boond’s MCP connector: the assistant has the signed-in user’s permissions, checked on every call. For any other tool, ask which account it connects with, then check that two people with different permissions do not get the same answer to the same request. If they do, the tool sees more than one of them is allowed to.

  • Can AI approve timesheets instead of the manager?

    Technically yes, since the MCP connector can approve or reject a timesheet. We recommend giving it the checking rather than the approval: the AI lists missing timesheets and inconsistencies, and the manager approves, because an approved timesheet triggers the invoice and commits the company to its client.

  • Does Balt replace BoondManager?

    No. Boond remains the system where requirements, consultants, timesheets and invoices live. Balt reads what it needs there, writes with your approval, and works from Teams, Slack or WhatsApp on everything that happens between Boond and the rest of your tools.

Sources

  1. Boond Help Center, MCP: Connect Boond to Your AI Assistant (October 2026)help.boondmanager.com
  2. Boond Help Center, BoondPulse: Your AI Assistant in Boond (October 2026)help.boondmanager.com
  3. Boond Help Center, Optimizing Your API Calls: Best Practices (October 2026)help.boondmanager.com
  4. Boond Help Center, OAuth2 Connection (April 2026)help.boondmanager.com
  5. Boond, external API documentationdoc.boondmanager.com
  6. Boond, New plans to grow with your company (April 2026, in French)boondmanager.com
  7. Boond, ERP and CRM for IT services and consulting companies (AI matching, skills files, bench control)boondmanager.com
  8. Boond, Balt partner pageboondmanager.com
  9. GitHub, silamir/boondmanager-mcp-server, a community MCP server for the BoondManager APIgithub.com
  10. Loopio, RFP Trends & Benchmarks Report 2025 (1,544 respondents)link.loopio.com
  11. Cobalt, Skills files: 7 mistakes that lose assignments (April 2026, in French)cobalt-ia.com

Balt does this work

Read next

Give Balt its first job.

Thirty seconds to add it to Teams and get to work.

Try for free