The chat that already knows your firm.
Not a back office with a chatbot bolted on top. It's a chat that has your projects, your people and your numbers in front of it, and that tells you when it can't answer instead of making something up.
An assistant, not an answer engine.
The difference isn't how it talks. It's what it can do while talking, and what it refuses to do.
Zero setup, zero training
The chat already knows your firm on day one. Getting started is five questions and an automatic import from Google Workspace: nothing to configure, and nothing to teach it.
Ask in plain language, not in dropdowns
"Find a Java consultant available from January for Milan on-site", instead of four filters and two clicks. And it speaks consulting: daily rate, assignment, bench, timesheet.
It has tools, not just words
It reads a project's margin, phase status, who's allocated at what capacity, recent activity, documents. And it drafts — a change request, a risk: drafts, doesn't save. Confirming stays with you.
It sees what you see
If you don't have access to a project's costs and margin, the chat won't comment on them: it says it can't see them and points you to who can. It doesn't estimate, round, or produce a plausible figure.
It speaks when it matters
Notifications fire only for real urgencies: an assignment expiring without renewal, someone stuck on the bench too long, a timesheet not submitted. Otherwise it waits until you open the chat.
@hice inside conversations
Colleagues chat inside hice, and the AI is summoned with @hice mid-conversation. No separate messaging tool to keep open.
If it can't do something, it says so
There's no polite answer for every question. When something is out of reach the chat declares it and sends you to support, instead of trying and quietly getting it wrong.
Drop a CV. AI does the rest.
Hice extracts skills, years of experience and seniority from the CV, then matches the profile against every open project and suggests the best fits. Zero copy-paste.
Candidate matching, explained all the way down.
A score without the why is a number nobody will actually decide on. Here the score comes last: first the criteria, the evidence and the doubts.
First it reads the request
The structured title is the requirement that rules; free text refines it. Every requirement lands in one of three levels, and doesn't change level for convenience.
Constraint
Declared mandatory. Missing is missing: no high score compensates for it.
Preference
Nice to have. It weighs on the result, but on its own it excludes nobody.
Inference
Implicit in the context: "migration" means experience on the legacy and the target, not just the target. It stays marked as an inference and never becomes a constraint.
If a requirement is ambiguous — "senior" without saying how many years — it isn't resolved arbitrarily: it's kept as an explicit threshold and the evidence speaks. The place to clarify is the chat, not the engine.
Then it reads the profile
Trajectory before titles
Seniority is estimated from role progression, project scale and the autonomy you can infer from the descriptions. A "junior" title with twenty consistent years is an inconsistency to flag, not an average to take.
When, and for how long
It doesn't count that a skill is on the list. A technology used eight years ago for two months isn't worth the same as one used last year for four years.
In what context
React in a game studio, in an e-commerce or in a keyword list are three different signals. Evidence is worth what the context it came from is worth.
Long lists don't convince
Many skills, few years and zero project detail is an explicitly negative signal, not a neutral one.
Aspiration isn't experience
"I'd like to work with X", courses without projects, and the job ad's requirements copied into the CV prove nothing.
When sources contradict each other
Seven different sources speak about a candidate, and they don't carry equal weight. The order is settled up front, not case by case.
A divergence between sources isn't averaged, it's declared. If the interviewer writes "elementary English" and the CV says "fluent", the criterion follows the highest and most recent source, and the report shows the conflict in the open with both citations.
- 1The structured interview assessment, the scorecard
- 2The notes of whoever ran the interview
- 3Verifiable work history: dates, roles, companies
- 4Certifications
- 5The CV text
- 6Self-declarations in screening forms
- 7The model's inferences, which never decide on their own
The report that comes out
Not a percentage and goodbye. Every piece of the result can be opened and challenged.
Breakdown by criterion
Every criterion with its state — supported, to verify, not supported, contradicted — and how much it weighed on the total.
How the score is built
The equation in the open. The number doesn't come down from above: you can reconstruct it.
Evidence considered
No verdict without a citation. A criterion is upheld or dismantled only by pointing at the extracts that prove it, and evidence quality is declared: high, medium or low.
AI profile analysis
Calibrated seniority, transferable skills, warning signs and interview questions, generated only on the real doubts the analysis surfaced. Never stock questions.
Comparative reading
Between two profiles it explains the trade-off rather than repeating the scores. The ranking is the deterministic engine's call: the comparison exists to explain it, not to rewrite it.
To verify
A criterion left unanswered is a legitimate outcome, and a better one than an invented certainty. Doubts don't get rounded away to make the number work.
You set the priorities
Every criterion sits in a priority tier and can be marked mandatory; two criteria in the same tier carry equal importance. Change the priorities and the shortlist recomputes. You can look at the same request from different matching perspectives, ask for five more profiles, go deeper on a single candidate, and dismiss a suggestion you don't want: dismissed, it doesn't come back.
And when the answer is no
If no profile covers a criterion, the report says so and points at what would widen the pool. If the analysis is partial, because some profiles haven't been judged yet, the status declares it instead of passing it off as complete. A weak pool stated plainly is worth more than groundless enthusiasm.
The chat is in the plan, not a separate module.
No extra charge for the AI, no usage meter to watch. Up to 10 employees, hice is free.