Which AI Agency Fits a Company Without Its Own IT Department?

"We want to use AI, but we have no IT department - who do you even turn to?" This question comes up strikingly often in initial consultations lately, and it has a sound reason: 79,000 IT positions are currently unfilled in Germany, the digital association Bitkom reports in its September 2026 labour market study. Of companies looking for IT staff, 97 percent struggle to fill positions and 31 percent receive practically no applications. For a mid-sized business this means: "just hire someone" is no answer - the position will most likely stay vacant.

That leaves the market. And the market is confusing, because "AI agency", "AI consulting" and "AI tools" sell three different things that look alike on websites. This article sorts the four provider types, shows which one works without an IT department, and names the checkpoints that let you place any offer in a first call.

Key takeaway: Without an IT department you need a provider that also operates the system - not just builds it. Four types compete: the toolkit (licence from about 9 to 60 euros a month) shifts building and maintenance onto you. Consulting (day rates of 1,200 to 2,500 euros) delivers a concept whose implementation stays with you. The agency with handover (fixed prices of 3,000 to 35,000 euros depending on integration) delivers a system you must run yourself. Only the agency with operation (setup plus a retainer of 2,000 to 8,000 euros a month, all figures from market overviews) contractually covers maintenance, monitoring and adjustment. For comparison: an in-house full-time position costs around 93,600 euros a year at the German statistical office's hourly labour cost of 45.00 euros - if you manage to fill it at all.

What is the difference between AI consulting and an AI agency?

Consulting analyses and recommends; an agency builds and operates. Consulting ends with a concept - potential analysis, prioritisation, roadmap - whose implementation stays with the customer. An agency ends with a running system.

The distinction sounds formal, but it is the most common misjudgement in first calls: people searching for "AI consulting" usually do not want to buy an assessment, they want a result. Consulting pays off when someone in-house carries the implementation or when a neutral second opinion is needed before choosing a vendor. Without an implementer it produces a document that is outdated after three months - while the actual work is still outstanding.

The practical consequence: buy the survey of your first process from the vendor who will later implement and operate it. With specialised providers this analysis is part of the offer, not a separate day-rate business beforehand. Split consulting bills the concept twice: once at the consultant, then again as process surveying at the implementer.

Which four provider types are on the market?

Four models exist side by side. They differ less in price than in a single question: who operates the system after day one?

The table also explains why price comparisons online are so confusing: almost always only the first line of the invoice is compared - licence or fixed price - while the decisive line is operation. A system nobody operates is not cheap, it is short-lived.

Why does handover fail without an IT department?

Because an AI agent is not a purchased appliance but ongoing operations. Three tasks persist after go-live, and each needs an owner:

First, the knowledge base. Your product range, prices and processes change; the agent only answers as up to date as its sources are maintained. Second, the rules: as soon as a process or a connected system changes - an ERP update is the normal case, not the exception - someone has to adjust handovers and exceptions. Third, the model layer: providers, model versions and prices change continuously; the right choice today is not necessarily the right choice in a year.

The figures back both sides. MIT Project NANDA found in 2025 that around 95 percent of the generative AI pilots examined produced no measurable return - almost always because of missing integration into process and data, i.e. precisely this ongoing work. And in the Bitkom survey of September 2026, 41 percent of AI-using companies name integration into existing systems as a major cost item, only 21 percent name licences. Operations is the work, not the purchase.

A handover without an internal operator therefore means: the system goes stale at the first process change, and the failure only becomes visible when a customer receives a wrong answer.

What must operation include so you need nothing yourself?

"Operation" is a stretchable term. For it to carry what it promises, six points belong in the contract, not in the presentation:

1. Hosting and availability with a stated server location, a data processing agreement and an availability commitment.
2. Model usage with clear billing - flat rate or by consumption, in the second case with a cap.
3. Knowledge base maintenance with a defined turnaround: how quickly does a new price list or a new process reach the system?
4. Rule maintenance when the process or connected systems change - the line item most often missing from quotes.
5. Monitoring and error analysis with a written response time, not "we will have a look".
6. The exit clause: configuration, knowledge base and documentation belong to you and are handed over in usable form at the end of the contract. Without this point your process is permanently attached to someone else's system.

What you still need in-house is manageable: a business contact who knows the process and grants approvals. Domain knowledge, not IT skills - the role can come from sales, accounting or service.

How do I recognise in a first call that the provider means operation?

By the answer to a single question: "What happens in the second month?" A provider who means operation describes concretely what arises then and who does it. Three warning signs show that the offer ends at go-live:

• "We document everything and train your team" as the only answer to every operations question. Training is handover by another name - it presumes someone who can be trained.
• No operating model in the quote. If the quote contains only one-off prices, operations is implicitly yours. Ask for the monthly item; if it is missing, the model is missing.
• No response time. "We are reachable" is not a commitment. A failure on Friday evening needs a written rule, not a promise.

Positive signs run the other way: a named contact instead of a ticket system, a short monthly report on volume, errors and changes, and a provider who raises the second-month question before you do. The complete list of points a first call should cover is in the checklist with twelve questions for an AI agency.

Does the agency's location matter?

In only one place: the survey. Measuring a process is faster on site - a site walk-through, an afternoon with the business department, a look at the filing system. In the ongoing operation of a managed system both sides work remotely anyway; what decides there is written response times and availability, not the postcode.

Anyone wanting regional availability should honestly class it as a comfort criterion, not a quality marker - and check in return whether a provider fakes proximity: whoever claims an office in your city while their imprint sits elsewhere has already answered the reliability question themselves. Transparent providers state their service area in the imprint and on their website instead of merely claiming a local presence.

What does the model cost - and what is the alternative?

The alternative to external operation is the internal position, and its cost is documented: the Federal Statistical Office puts labour costs per hour worked in Germany at 45.00 euros for 2025. A full-time position thus costs around 93,600 euros a year (45.00 euros x 40 hours x 52 weeks, counted with 52 weeks because payroll costs continue through holiday and sick weeks) - plus recruiting, onboarding and the outage risk when precisely that one person resigns. And this under the conditions of the Bitkom figures: 97 percent of companies searching have staffing problems.

The retainer for operation sits at 2,000 to 8,000 euros a month, i.e. 24,000 to 96,000 euros a year (market overviews) - for a managed function instead of a person, without a search process and without holiday cover. For a first bounded process the bill sits at the lower end; only when several processes are operated in parallel does the retainer approach the staff calculation.

Honesty also means the calculation runs the other way: anyone planning to build a digitalisation position anyway is better served by handover with an internal operator - external operation is the bridge for the years in which that position does not exist or cannot be filled. What setup, model usage and the remaining cost types comprise is calculated in the article on the cost of AI automation; where the line between toolkit and custom development runs is covered in the comparison of standard chatbot versus AI agent.

Frequently asked questions

We have no IT department. Can we still use AI? Yes, and without hiring a specialist, provided the vendor contractually takes over ongoing operation: hosting, model usage, maintenance of the knowledge base and rules, monitoring and response times. What you need in-house is a business contact who knows the process and grants approvals - not IT skills. It only becomes problematic with the handover model, where operation sits with you after go-live.

Is AI consulting enough when we are just starting out? As a first step, yes; as a solution, no. Consulting delivers analysis, prioritisation and a roadmap - it builds and operates nothing. If you lack implementation capacity, buy the consulting from the same vendor that will implement later, otherwise you pay for the concept twice: once at the consultant and again as the survey at the implementer. With specialised providers, surveying the first process is part of the offer.

What is the difference between handover and operation? With a handover you receive a finished system plus documentation and carry everything from that day on: maintaining the knowledge base, adjusting rules, monitoring, error analysis. With operation, exactly that stays with the vendor and is billed monthly. Handover is the right choice when someone in-house can carry operation professionally and in terms of time - without that person the system goes stale at the first process change.

How do I recognise that a quote silently assumes an IT department? In three places: the quote ends at go-live instead of with an operating model, training and documentation are given as the answer to every operations question, and no line item covers maintaining the knowledge base and rules. Ask concretely: who updates the knowledge base when our product range changes, and within what response time? The answer "you can maintain that yourself in the system" means: you.

What does ongoing agency support cost per month? Market overviews cite retainers of 2,000 to 8,000 euros a month for operation, maintenance and further development, depending on the number of processes and systems covered. A single bounded use case usually sits at the lower end. More important than the amount is the split: hosting and tool licence, usage-based model costs with a cap, and the maintenance share should each be listed separately, otherwise quotes cannot be compared.

Does the agency have to be located in my region? No. With a continuously operated system, availability and contractually agreed response times decide, not the postcode. On site, proximity helps during the survey - a site walk-through, conversations with the business department - but in operation both sides work remotely anyway. Anyone wanting regional availability should treat it as a comfort criterion, not a quality marker.

---

What a first engagement with survey, fixed price and subsequent operation looks like concretely is described on the page about AI process automation for companies. This is not legal advice; figures on third-party providers and market ranges come from the sources linked below with the dates stated there - only the individual quote is binding.

Sources

• Bitkom (3 September 2026): IT-Fachkräftemangel hat sich in den letzten drei Jahren halbiert - 79,000 unfilled IT positions, 97 percent with staffing difficulties, 31 percent without applications; survey of 853 and 602 companies with 3 or more employees
• Bitkom (14 September 2026): Erstmals nutzt die Mehrheit der Unternehmen KI - including the cost blocks integration (41 percent) and licences (21 percent), surveyed in weeks 28 to 33 of 2026
• Federal Statistical Office (29 April 2026): One hour of labour cost an average of 45.00 euros in 2025
• n8n: Pricing, Zapier: Pricing, Make: Pricing - list prices of the toolkit models, as of September 2026
• MIT Project NANDA (2025), reported by Fortune: MIT report: 95% of generative AI pilots at companies are failing
• Market ranges for day rates, fixed prices and support retainers: Jean Hinz, KI-Agentur Kosten and Codapp, Was kostet eine KI-Agentur - provider overviews, as of August/September 2026