Most AI projects stall for reasons that have nothing to do with technology. The model connects in a day, RAG is configured in a week — and then the project sits for three months because nobody in the company can say where the current price list lives or who signs off that an answer is correct. At SUDUS·TEAM we have walked this road often enough to write down a schedule that actually holds.
Below is a 30-day plan: what happens each week, what the contractor does and what you will have to do yourself, with honest markers at the points where things usually stop. The plan assumes a single-purpose AI agent — that is the scope that genuinely fits into a month.
Before the start: three decisions, without which do not begin
Decision 1. One process, not "AI for the company"
The most common cause of failure is a blurred goal. "We want to introduce AI" is not a task. "We want an agent that answers questions about order status, delivery and returns without a manager" is a task.
Pick a process that is at once repetitive, high-volume (200+ cases a month), governed by clear rules, and free of legal responsibility for the decision.
Decision 2. A success metric expressed as a number
Agree in advance which number means the project worked. For example: "70% of enquiries close without a manager, at a customer rating of at least 4 out of 5". Without it, a month from now you will be arguing about impressions.
Decision 3. A project owner on your side
You need one named person who knows the subject, has 4–6 hours a week and the authority to settle disputed points. Not the director "in general", not "someone in support". Projects without this role are the ones that collapse.
Week 1: audit and data preparation
What the contractor does
- Breaks the chosen process into steps and reads real conversations from recent months
- Classifies enquiries by type and frequency — it becomes visible which 20% of questions create 80% of the load
- Designs the agent's persona: tone, limits of competence, when to hand over to a person
- Draws up the list of integrations needed
What you do
- Export 200–500 real customer enquiries
- Gather the documents for the knowledge base: prices, terms, policies, FAQ
- Remove contradictions — if a price differs across three files, decide which one is right
- Provide test access to the systems the agent has to work with
This is where most projects stop. "We will collect the documents next week" gets repeated four times in a row. Plan for 10–40 hours of your team's work and appoint someone responsible. Exactly what to collect and in what form is set out in RAG explained.
Week 2: a prototype in the sandbox
What the contractor does
- Indexes the knowledge base and configures hybrid search
- Writes the system prompt and the persona
- Connects the first model and assembles a working prototype in a closed environment
- Prepares a control set: 50–100 questions with known correct answers
What you do
- Read the agent's first answers and mark where it is wrong
- Refine the tone: too formal, too familiar, too wordy
- State what the agent must never say
By the end of week two there should be a prototype you can talk to. It still gets things wrong, but you can already see whether the task is solvable at all. If the quality is frankly poor at this stage, it is better to stop and revisit the brief than to carry on.
Week 3: integrations and guardrails
What the contractor does
- Connects the agent to live systems: website, messenger, CRM, mailbox
- Implements the tool functions: check status, create an enquiry, book a slot
- Sets up guardrails: protection against prompt manipulation, a ban on disclosing other people's data, cost limits
- Writes the handover path for difficult cases to a human manager
- Deploys the dashboard: conversations, ratings, errors, spend
What you do
- Provide production access and agree the permissions
- Define the conditions under which the agent is obliged to call a human
- Prepare the support team for the fact that some enquiries will now arrive already worked through
This is where the technical surprises surface: the CRM has no field for what you need, the API returns data in an odd format, the legacy accounting system has no documentation. Build in a buffer — week three is where deadlines most often slip.
Week 4: testing on people and launch
What the contractor does
- Runs the control set and measures accuracy — the target is 85% or above
- Checks behaviour on questions outside the knowledge base: the agent must honestly say "I don't know"
- Iterates on the prompt and the chunking based on the results
- Optimises token cost: model routing, caching, history limits
- Launches on part of the traffic first
What you do
- Give the agent to real in-house experts for two or three days
- Collect their comments in a structured list, not as one stream in a chat
- Make the launch decision against the agreed metric
A proper launch is gradual
Do not switch the agent onto 100% of enquiries at once. The working pattern: first it answers only outside business hours, then on 30% of daytime traffic, then fully. Each stage runs 3–5 days with a review of the logs. That way mistakes stay cheap.
What happens after day 30
Launch is not the finish line. The first month or two decide whether the agent becomes a working tool:
- A weekly look at the logs. You see where the agent handed over something it could have closed, and where, conversely, it should not have tried.
- Topping up the knowledge base. Every unanswered question is a reason to write a document.
- Cost control. The real token bill compared against the forecast. The arithmetic is in how much an AI agent really costs.
- Extending to adjacent tasks. Once the first process is stable, the second agent takes half the time — the knowledge base and the integrations already exist.
Readiness checklist
Go through this before signing anything. Every "no" costs you either days on the timeline or quality in the result.
- One specific process has been chosen, not "AI in general"
- That process produces at least 200 similar cases a month
- A success metric as a number is written down, along with who measures it
- A project owner with 4–6 hours a week has been appointed
- There is an export of real enquiries from the last 3–6 months
- The current price list and terms exist as text, not as images
- Contradictions between documents have been resolved — one fact, one source
- It is known which systems the agent integrates with and who grants access
- It is agreed when the agent must call a human
- The support team has been warned that the shape of the enquiry flow will change
Eight out of ten "yes" and you can start calmly. Fewer than six and you need a week of preparation first, otherwise you will be paying the contractor to wait.
The usual rakes to step on
"Show us the finished thing" approvals
The client does not want to look at intermediate versions and waits for the final one. In week four it emerges that the tone is wrong and half the answers contradict company policy. Reworking eats another fortnight. Look at the prototype in week two, however raw it is.
"Let the agent answer absolutely everything"
The temptation to build a universal assistant on the first attempt. In practice this dilutes the knowledge base, lowers accuracy on the core questions and makes the errors unpredictable. A narrow agent that does one thing well is worth more than a broad one nobody trusts.
No access to real conversations
"We roughly know what customers ask" does not work. Management's sense of question frequency is usually off by a factor of two. An export of the actual correspondence is the single most valuable material in the whole project.
Launching on all traffic at once
Tempting, because "why wait". But the first days always bring surprises, and it is better that 30% of users see them rather than everyone. A gradual rollout costs nothing and protects the reputation.
Nobody looks at the logs after launch
The project is delivered, everyone exhales, nobody opens the dashboard. Three months later it turns out the agent is handing half its conversations to managers because of one question that is missing from the base. Five minutes a week in the logs pays back many times over.
A realistic schedule: when 30 days will not happen
Honestly about timelines. A single-purpose agent fits into a month provided the data is ready and the access exists. It will not fit for:
- A multi-agent system — 35–60 days, because orchestration and quality control between agents are added
- Projects with document recognition — add 2–3 weeks for working with scans and photographs
- Companies that approve across several departments — the timeline is set by the meeting calendar, not by development
- Local model deployment — add a month for infrastructure and security
How to measure the result: five metrics
So that a conversation about effectiveness a quarter from now rests on numbers rather than impressions, record the baselines before launch and compare against them.
1. Share of enquiries closed without a human
The headline metric. Counted as the percentage of conversations after which the customer did not contact a manager within 24 hours. A realistic target for month one is 55–70%; after a quarter of work on the knowledge base, 75–85%.
2. Accuracy on the control set
The same set of 50–100 questions is run monthly. A drop in accuracy is an early signal that the base has gone stale. Below 85% is a reason to stop and investigate.
3. Time to first response
This is where the change is most visible: instead of 40 minutes during working hours and a night's wait outside them, seconds around the clock. This metric explains the project's value to people outside IT better than any other.
4. Customer rating
A simple two-button form at the end of a conversation. What matters is not the absolute figure but the trend and the text of the negative ratings — they show what the knowledge base is missing.
5. Cost per closed enquiry
Token and maintenance spend divided by the number of conversations closed without a human. You can put that number next to the cost of an hour of a manager's time, and that is what makes the ROI conversation concrete.
All five are collected automatically on the dashboard we hand over with the agent. Look at them weekly in the first month and monthly after that.
Frequently asked questions
How much of our team's time will the project take?
Roughly 30–50 hours over the month, most of it in week one. Mainly gathering and verifying documents plus reading the agent's answers. It splits between the project owner and one or two subject experts.
Can it be launched faster than 30 days?
In 10–14 days, if the task is narrowed to answering the top 20 questions with no external integrations. That is a good way to test the hypothesis cheaply and expand afterwards.
What do we do if quality drops after launch?
The most common cause is a stale knowledge base or a new type of question appearing. You look at the logs of enquiries the agent did not close and top the base up. The second most common case is someone switching to a cheaper model without retuning the prompt.
Who is responsible if the agent makes a mistake in front of a customer?
Responsibility always sits with the company, which is exactly why guardrails and a gradual launch exist. Technically the risk is reduced by forbidding the agent to speak outside its competence, mandating handover on sensitive topics, and logging every conversation for review.
Do we need a dedicated person to maintain the agent?
For one agent, no — 2–4 hours a month from a content manager plus the contractor's support is enough. A dedicated role appears at around five agents across different departments.
In short
Thirty days is a realistic timeline for a first AI agent, but only if the company arrives with its data ready, one clear task and a person accountable for it. The technical side is rarely the bottleneck; the bottleneck is usually not being able to answer "where is our current price list".
Want to do this with a contractor who names the risks at the first meeting rather than in month three? Write to us. A walkthrough with concrete numbers for your process takes 30 minutes and costs nothing.