What AI actually changes at work.
Concrete takes on AI agents in small and mid-sized businesses: what they do, what they cost, what to hand them. No empty promises.
ERP or CRM for an IT consulting firm: which one do you need?
A CRM is enough for an IT consulting firm while each fact lives in one place. You need an ERP once availability, rates and timesheets are copied between tools.
AI agent: what it is, and how a small business should choose one
An AI agent takes a goal and acts in your tools to reach it, where an assistant only writes. A definition, selection criteria and a trial for small businesses.
Answering an IT services tender: what actually wins
An IT services tender is won before the writing starts: an honest go/no-go, available consultants, a price read off the grid. Why most responses lose.
LinkedIn boolean search: the strings that actually find people
LinkedIn boolean search accepts AND, OR, NOT, quotes and parentheses, with no wildcard. Ready strings for Java and DevOps roles, and what they miss.
ChatGPT for recruiters: what it does well, and where it stops
ChatGPT drafts a job ad, a Boolean string or an interview guide in a minute. It sees neither your ATS nor your clients, and a pasted CV leaves your records.
Setting an IT consultant’s day rate, and negotiating it
An IT consultant’s day rate is built from daily cost and a target margin, checked against the 2026 market, and negotiated with a floor written down beforehand.
A consultant’s trial period in a French IT firm: length and exit
Four months for a Syntec manager, renewable once, two or three for other staff. Ending a trial when a mission falls through risks an abusive termination claim.
The non-solicitation clause in IT consulting: what it protects
A non-solicitation clause stops the client hiring the consultant. It binds two firms, not the employee, and since 2021 must be proportionate under French law.
The cost of a bad hire: the honest calculation
The “30% of first-year salary” has no verifiable source. For an IT services consultant, the cost of a bad hire is counted in days of lost margin.
Employee referral bonus: how much, and why it is not enough
A referral bonus averages €774 in France and is taxed as salary. What makes employees refer has less to do with the amount than with being involved.
Interview scoring grid: building a structured interview
A structured interview with a scoring grid predicts performance twice as well as a free-form one (0.42 against 0.19). Method, rating scale and legal limits.
Writing an IT job ad that developers actually answer
An IT job ad that attracts states the pay, stack, remote days and contract in under 300 words. The EU pay transparency directive makes the range unavoidable.
Getting on an enterprise supplier list costs before it pays
A place on an enterprise supplier list guarantees no work. The paperwork is paid every six months; orders depend on a relationship the panel does not create.
Candidate data retention: the three clocks that apply
Two years in a talent pool after the last contact, five years in an archive for evidence. What an AI agent remembers about a candidate runs on the same clocks.
One tool that writes code: why our agent has nothing else
A tool the model can see is paid for at every step. A function that code imports costs nothing until it is used, and the gap grows with every integration.
The intake meeting decides the placement before the first CV
A rushed intake meeting cannot be fixed by sourcing. Seven questions qualify a client’s need, and the last one decides whether to take the role at all.
In IT consulting, the margin is lost at renewal
A day rate renewed without review loses about a sixth of its margin over three years of 3% raises. Renegotiation is a date to keep, not a talent.
The reference check nobody really does
The reference check as usually run predicts almost nothing. Structured around verifiable facts, it becomes a real control again.
The six-second CV scan: what the study actually measured
The figure comes from a 2012 study of 30 recruiters, published by a company selling CV rewrites, and it measured a first screen, not a read.
A consultant’s skills dossier is a sales document
A skills dossier is written by the firm, for one client and one assignment. You may select, order and rephrase, never add. And who should sign it off.
The talent pool nobody searches before paying to source
Close to half of sourced hires come from profiles already in the database, and most teams pay for external sourcing before they have looked there.
When a recruiter leaves, their pipeline leaves with them
A departure costs 13 to 17% of the affected accounts’ sales the following year. What leaves is not the database: it is the promised follow-ups and the context.
AI automations in an agency: the ones that return nothing
An automation returns nothing until the time it saves has a written destination. Three tests before you install the first one.
How many roles do you work for free?
A 45% fill rate means more than one role in two will never be invoiced. That is the real price of contingent work.
Should you let an AI agent drive your browser?
An agent inside a logged-in browser inherits every one of your sessions at once. The Comet case, the 2026 numbers, and the boundary to hold.
MCP and recruitment: what the protocol does not solve
MCP solves access to systems you own. An agency’s problem is its client’s ATS, and the protocol changes nothing there.
Nobody watches the agent that runs at 3 a.m.
A scheduled run has no requester. It inherits the rights of whoever set it up, outlives their departure, and fails in silence.
AI workers: a digital colleague is onboarded, not installed
The colleague analogy holds on four points and breaks on two. The two breaks decide the outcome of the first thirty days.
The consultant nobody has called in six months
A consultant on assignment has two employers and only one of them talks to them. IT services attrition is a calendar problem before it is a pay problem.
The counter-offer statistic everyone quotes is wrong
The “80 to 90% leave within a year” traces to no study. The figures that do have a source say 32%, and that changes your method.
A meeting recap is not a follow-up
75% of professionals use an AI notetaker, only 43% inside large enterprises. What the recap does not solve, and where the delay really sits.
What an AI agent needs from your ATS and CRM
Five properties decide whether an agent can work inside your ATS, and none of them is “has AI built in”. The list, in the order they break.
An ATS, a CRM and a spreadsheet: which one is right?
The right metric is not how many tools you run, it is how many places the same fact is written. One rule, and what it costs.
Your AI agent has an expiry date
Claude Sonnet 4 and Opus 4 were retired on 15 June 2026 after 62 days’ notice. What a model migration actually breaks, and how to prepare for it.
Your revenue comes from clients you never call
79% of agencies make at least half their revenue from existing clients, and almost all of their commercial effort goes elsewhere.
Agent washing: telling a real agent from a chatbot
Gartner estimates only around 130 vendors, out of thousands, sell a real agent. Five questions to settle the category before you buy.
An AI agent on Teams or on Slack: what changes
Teams holds about 37% of the market against 13% for Slack, yet the agent ecosystem is built on Slack. What each surface forbids, and what it costs.
The MCP server you plug in is a supplier
An MCP server holds your credentials and feeds your model text it reads as an instruction. The 2026 measurements, and three questions to ask first.
Claude Tag: what an agent living in Slack changes
Claude Tag installs an agent as a member of a Slack channel, with an ambient mode that speaks unprompted. What it changes, and what it does not do.
Cobalt and Balt: two products, and which one you need
Cobalt is the ATS-CRM where your data is written, Balt the agent that works in Teams and Slack. Balt runs without Cobalt, and here is when.
Entry-level hiring is collapsing, and the pyramid with it
The employment gap for 22-25 year olds in AI-exposed work has reached 19%. This is not a redundancy plan, it is a hire that never happens.
Running two agents that both know your company
Yes, on one condition: their contexts must not overlap. And to cooperate, one of them has to stop being a colleague for the duration of the call.
Specialised or generalist agents: which will win?
The question is badly posed. What protects a vertical agent is not its knowledge, which the model will eventually have, but the responsibility it accepts.
Must you tell a candidate they are talking to an AI?
Yes if they converse with the agent, no if a human reviews and sends. Article 50 has applied since 2 August 2026 and the line is not where you think.
Time-to-fill: 18 days for one firm, 81 for another
The top decile places in 18 days, the bottom in 81, in the same market. A 4.5x gap does not measure scarcity, it measures dead time.
Uniform governance is what makes AI agents fail
Gartner expects 40% of enterprises to demote an autonomous agent by 2027. The cause is one blanket policy applied to agents of unequal reach.
Your AI agents have no identity, they have your keys
An agent does not have an identity problem but a delegation one: it acts with its rights instead of yours. The test that reveals it fits in one question.
Your agent will subcontract to other agents
Yes, and it is already standardised: over 150 organisations behind A2A in one year. An expert agent is called like a tool, provided it keeps nothing about you.
An AI agent’s memory: what it has to forget
A memory that swells unfiltered drops accuracy from 39 to 13 %. What an agent forgets matters as much as what it keeps, and that is settled at design time.
Can you automate 100% of a company?
No, and the limit is not model capability. A company is a person that answers for its acts: automation stops exactly where somebody has to answer.
Cited without being visited
Between 82 and 88% of citations in an AI answer produce no visit at all. The click stops being the unit of measure, and recommendation replaces it.
How to test an AI agent before you sign
With thirty cases from your own history, whose right answer you already know, replayed three times each. A demo proves none of that.
One candidate profile in four will be fake by 2028
Gartner expects one profile in four to be fake by 2028, and half of applications to remote IT roles. 41 % of companies have already hired one.
Two thirds of your team use a banned AI
66 % of office professionals used an AI they believed was banned, and over a third entered client data into it. Banning harder only moves the usage.
Skills or subagents: what actually works
A single agent with well-written skills matches multi-agent architectures. But an auto-generated skill adds nothing, and a bad one degrades performance.
The AI doom loop in recruiting
Candidates apply with AI, recruiters screen with AI, and response rates collapse on both sides. Getting out means changing channel, not adding tools.
The AI match score is the wrong instrument
A score of 87 % cannot be argued with, only applied. 75 % of companies let an AI reject without review, and 26 % of candidates call that fair.
Intelligence is becoming free. Where does the value go?
The cost of the same level of answer fell 280-fold in eighteen months. When intelligence costs nothing, what becomes scarce is context and the right to act.
The end of interfaces, not the end of software
At Supabase, 60% of new databases are launched by an agent, and probably 90%. What is dying is not the software: it is the access layer.
The four eras of AI, and the one now starting
Symbolic, machine learning, deep learning, generative then agentic. What comes next is being decided on memory and world models.
Twelve agents that do not talk to each other
A company runs twelve agents on average and half of them work alone. The right criterion is not how many you have, it is how many contexts you duplicate.
What is left to humans when agents do the work
It is not the job that disappears, it is the pyramid. The volume work that trained juniors is going, and nobody has planned a replacement.
Why an AI agent costs 30 times a chatbot
Because an agent rereads its whole context at every step. Resent context is 62 % of the bill, and past fifty steps the multiplier goes beyond thirty.
The work that looks finished and is not
40 % of workers get it from a colleague, and each instance costs almost two hours to whoever picks it up. The cost does not vanish, it changes person.
AI Act deferred to 2027: what still applies
The high-risk rules move to 2 December 2027. But telling a candidate they are talking to an AI has been required since 2 August 2026, and GDPR never moved.
AI agent or automation: which one for which task
An automation follows rules, an agent makes decisions. To choose: if you can write the script, automate. If you can only describe the outcome, delegate.
Training your teams on AI is now an obligation
Article 4 of the AI Act has required a sufficient level of AI literacy since February 2025. It is an obligation of result, with no format imposed.
Automating LinkedIn in 2026: what it really costs
40% of accounts using non-compliant tools were restricted in Q1 2026. And what you are risking is not a company account.
Bench time is not a sourcing problem
An 11.4% average bench rate, 41 days, €5,480 a month per consultant. The cost is already committed once you learn about the roll-off: the lever is the date.
Candidate ghosting is a capacity problem
53% of candidates were ghosted in 2026, up from 38% in 2024. It is not rudeness: it is a saturated queue, and AI is the cause before it is the cure.
Can a CV manipulate an AI agent?
Yes. Text hidden in a CV gives an order to the agent reading it. It is the least covered flaw in enterprise agents, and no filter fixes it.
Delegating to an AI agent: where to start
Start with work you can check in ten seconds: follow-ups, interview notes, record updates. Here is the order that works, week by week.
Does a recruitment AI agent require a DPIA?
Almost always yes. France’s CNIL names algorithmic candidate screening explicitly, and a missing DPIA is a priority enforcement theme for 2026.
Before evaluating an AI agent, look at your data
A candidate database loses about 30% of its accuracy a year. An agent reasons soundly over that wrong data, confidently, which costs more than no agent.
Should you require your AI agent to speak MCP?
MCP is genuinely good news, but it is not a buying criterion. It is an exit criterion: what matters is the day you change tools.
What an AI agent costs, and why not per seat
Between €200 and €3,000 a month recurring. But the real question is the billing unit: per-seat pricing is structurally broken for an agent.
What an AI agent should not be able to do
A good agent refuses. Deciding for you, acting outside its scope, silently retrying: five things it should be incapable of, by design.
Who approves what when an AI writes to your candidates
No, approving everything is the same as delegating nothing. The rule that holds: approval whenever a message leaves the company or closes a door.
Why 40% of AI agent projects will be abandoned
Gartner expects over 40% to be abandoned before the end of 2027, for three reasons: no measurable ROI, no governance, no integration. All three are preventable.
Must you consult the works council before deploying AI?
In France, yes. A court suspended two HR tools in January 2026, an appeal court two more in May, with daily penalties, and no need to prove urgency.
No article in this category yet.

