AI assistants · in the working day, not in an extra tab
An assistant somebody still uses after twelve weeks.
CoderConda sets up AI assistants for companies that sit in the tool that is open anyway, know your corpus and admit when they do not know something. The technology is the easy part; in the working day these three things are what count.
- Teams
- Outlook
- Intranet
- 07:52State of the variations, from the ERP
- 10:15Travel expense rule, with the clause
- 13:40Rejection draft, in the approval folder
- 15:05No entry — recorded as a gap
- 16:30Handover signature, with the date
- Tally1 draft, 0 invented answers
An ordinary Thursday · the gap is the part that gets better next month
01 — The problem
The assistant does not get switched off.It gets forgotten.
There is rarely a decision against it. Usage simply falls, because three answers in a row were worth nothing and nobody tries a fourth time. The licence gets paid all the same.
- Week 1
The rollout goes well. Everybody tries it, usage is high, there is an email to all staff.
- Week 3
The first real question gets answered wrongly, because the policy sits in the intranet in three versions.
- Week 6
The assistant lives in its own tab. Anyone in a meeting keeps asking the colleague next to them.
- Week 12
Usage is down to four people. The licence runs on, because it is in the annual contract.
02 — The ladder
Four rungs.It may stop on any of them.
An assistant that tries everything itself is the one nobody believes after four weeks. The useful one knows its rung and climbs one higher as soon as its own does not suffice, up to the rung where a human stands.
Rung 1 · Answering
It knows, and says where from
“From 30 km the away allowance applies. Travel expense policy 2025, clause 3.”
The answer comes with its source and date so it stays checkable.
An answer from the corpus
Rung 2 · Searching
It does not know at once, so it searches
“The manual says nothing about it. But the DMS holds a record from 14 June that answers the question.”
It searches file store, inbox, ERP and tickets, always with the asker’s permissions.
Search across every system
Rung 3 · Preparing
It prepares, but does not decide
“The draft is in the approval folder, with the three sources underneath. Nothing goes out without you.”
From here on it is an agent, and from here on limits live in code.
A draft, not an action
Rung 4 · Handing over
It hands over to a human
“Nothing in the corpus covers this. I have passed the question to legal and attached your number.”
It says who it is passing to and what it has already gathered.
The most important rung
03 — What gets built
One knows the world.The other knows your company.
An assistant that comes with a licence can write, summarise and translate, and do it well. It just does not know your corpus, and that is exactly what people ask about internally.
| Part | Off-the-shelf | Your own assistant |
|---|---|---|
| CorpusMost of the battle | General knowledge. | Your corpus, with the source. |
| PlaceIts own tab costs half the usage | In a window of its own. | In Teams, Outlook, the intranet. |
| Permissions and contextRow level security | Everyone sees the same. | Those of the signed-in colleague. |
| Limits and gapsIn code, not in the prompt | Phrases something plausible. | Says so and records the question. |
| RolloutNot technology | Buy a licence. | Eight weeks with one department. |
| Monthly reportNot technology | Usage figures. | Also the list of open questions. |
Six parts, and two of them are not technology.
04 — The self-check
Five statementsto settle before the technology.
Tick what is true for you today. The reading is the same one I would give in a first conversation, even when I advise against an assistant because something else is missing first.
All three outcomes stay in view
- 0 to 1 of 5
The corpus first, then the assistant.
At this stage I advise against an assistant and start with a knowledge base instead. Otherwise you build an interface over a corpus that cannot answer the questions.
Next step: survey the corpus
- 2 to 3 of 5
Possible, with preparation.
The groundwork is partly there. Before building I settle the missing points, which takes one or two weeks, not months.
Next step: close the gaps, then scope
- 4 to 5 of 5
A good moment for one department.
Corpus, questions, ownership and place are in hand. I start with one department, and you decide about expanding after eight weeks, based on usage.
Next step: scope for one department
05 — Limits & operations
Six commitmentsbefore the first use.
- 01
Sources on every answer
Without a source an answer cannot be checked, and what cannot be checked stops being believed after the first mistake.
- 02
Not staff monitoring
Analysis happens at the level of topics, not people. Put in writing before the first colleague types anything.
- 03
The asker’s permissions
The assistant never searches with its own permissions. Otherwise it would be the easiest route to documents somebody may not open.
- 04
What it does not do
Legal advice, employment-law assessments, binding commitments and anything that gets signed. Questions like that get passed on, not answered.
- 05
Servers and data
Operated in Germany, GDPR, a data processing agreement. On request entirely inside your own network, with a model that never leaves the building.
- 06
Stopping is provided for
If the numbers do not hold after eight weeks, I say so. You get the corpus, the analytics and the documentation as an export, and the contract ends.
06 — Plain answers
Six questionsthat come before the call.
01What is the difference between an assistant and a chatbot?
A chatbot answers questions from outside using public content. An assistant works inwards: it knows the signed-in colleague, sees only their areas and answers from the internal corpus with a source. That sounds like a small difference and it is the whole difference where permissions are concerned.
02Why does nobody use assistants like this after a few weeks?
Usually for three reasons. It sits in its own tab instead of the tool that is open anyway. It does not know the internal corpus and answers generically. And it never says that it does not know. After the third useless answer the trust is gone, and it does not come back by itself.
03Do I have to tidy up my documents first?
Not beforehand and not completely. The sensible route is the other way round: start with the corpus you have and use the list of unanswered questions as the tidy-up plan. Nobody else writes that list for you, and after four weeks it is more concrete than any concept paper.
04What about the works council?
An assistant that takes questions from employees is subject to co-determination, and that belongs at the start. What gets logged, who sees the analytics and what is explicitly not analysed is something I put in writing before the first colleague types anything.
05Can I start with one department?
That is the better way. One department, one corpus, one place in the working day, eight weeks. After that, usage figures and the list of open questions show whether expanding is worth it. A rollout to everybody before that measurement is the most expensive way to find out.
06Does this work without the cloud?
Yes. For sensitive corpora the model runs on your hardware and no question leaves the building. That costs compute and some answer quality against the large models. What that means in your specific case I measure against your question set rather than assert.
07 — Appointment
Which departmentgoes first at your company?
Name the department and the five questions asked most often inside it. From those I can see whether the corpus carries, and I will tell you if something else should come first.
The calendar is hosted by cal.com.
Loading it opens a connection from your browser to cal.com (USA), which transmits your IP address; cal.com may set cookies when it loads. You can also go straight there instead.
The group
- /19RAG & knowledge baseAnswers from your own corpus, with the passage they came from.
- /20Enterprise searchOne search across every system, with the searcher’s own permissions.
- /21AI agentsAn agent gets tools, a task and a limit.
- /22MCP developmentConnect tools once instead of once per client.
- /23AI assistantsYou are here
Five services on one foundation: a corpus, a search, permissions on the row.