AI strategy for businesses: a step-by-step guide
A practical step-by-step guide to putting together an AI strategy: map workflows, choose the right use cases, build the data foundation, choose an approach, run a pilot and bring employees along, plus when you should not automate.
Table of contents
An AI strategy for businesses is the prioritised plan for where AI creates value, how it gets into operation, who owns it, and how you measure the effect, not a list of licences.
An AI strategy doesn't need to be a 40-page document from a consultancy. For most businesses it comes down to six concrete steps: map your workflows and find the bottlenecks, choose two or three use cases that are repetitive and rule-based, make sure your data and systems talk to each other, choose an approach (ready-made AI software or an AI agent in your own instance), run a limited pilot with measurable success criteria, and bring employees on board from day one. The rest of this article walks through each step in practice, including the section most guides skip: when you should not automate.
AI strategy is a plan for where and how your business uses artificial intelligence to solve concrete business problems: which workflows are automated, which data is required, who is responsible, and how you measure whether it works. In other words, it is not a technology list but a prioritised sequence of decisions.
The need isn't theoretical. According to Statistics Denmark, the share of Danish companies with at least 10 employees that use artificial intelligence rose from 15% in 2023 to 28% in 2024: almost a doubling in a single year. That also means that the gap between companies that work with AI in a structured way and companies that buy licences blindly will show up in the accounts within a few years.
What is an AI strategy, and what isn't it?
An AI strategy answers four questions: Where in the business does AI create value? How does that value get into operation (tools, data, integrations)? Who owns the initiatives and decides when something needs adjusting? And when do you know whether it works, that is, which figures do you measure?
It sounds banal, but by far the majority of "AI strategies" in Danish companies are in reality shopping lists: a handful of licences for generative tools, with no link to workflows, data or goals. That isn't a strategy, it's an experiment without a hypothesis.
Strategy is prioritisation, not technology
The core of a good AI strategy is deciding what to leave out. You can't automate everything at once, and you shouldn't try. The strategy is the order in which you tackle things: first the workflows where the gain is greatest and the risk smallest, then the harder ones. The technology choice (which AI software, which model, which platform) comes after the prioritisation, not before.
AI automation vs. classic automation
Classic workflow automation follows fixed rules: "when X happens, do Y". Companies have done that for years with spreadsheet macros and integration tools. AI automation adds a layer on top: the system can read unstructured text, assess content, summarise, classify and draft, tasks that previously required a human. According to Statistics Denmark, the most widespread AI applications in 2024 are precisely text production (18% of companies) and text analysis (17%), while around one in ten companies use AI for workflow automation and machine learning for data analysis. That combination (rule-based automation plus language understanding) is what makes the next few years' wave different from the previous one.
Step 1: Map workflows and bottlenecks
Before you talk about AI at all, you need to know what your employees actually spend their time on. Not what the organisation chart says: what really happens between 8 am and 4 pm.
How to map in practice
Choose two or three teams (e.g. sales, customer service, finance) and run a simple exercise: for one week, each employee notes the tasks that are repetitive (happen daily/weekly), manual (require clicking, copying, rewriting) and annoying (what people grumble about by the coffee machine). You don't need process diagrams to begin with: a list with three columns is enough: the task, how often it is carried out, and how long it takes.
Then look for three patterns:
- Seams between systems. Places where data is moved by hand from one system to another: from the inbox to the CRM, from the webshop to the accounting system, from one spreadsheet to another. These are classic candidates for workflow automation.
- Waiting for people. Tasks that sit idle because they are waiting for one particular person to read, approve or forward them. The bottleneck is rarely the work itself, it's the queue.
- Tasks nobody owns. Things that "just get done" by whoever happens to notice them. They are often the first to slip when things get busy, and the best candidates for being put into a system.
Convert the bottlenecks into hours
Once the list is made, prioritise it using a rough calculation: how many hours a month does the task cost across the team? You shouldn't spend a week on precise measurements; an estimate at the level of hours is enough to rank them. The point of step 1 isn't perfection; it's to replace gut feelings ("we spend far too much time on quotes") with a concrete, prioritised list you can choose from in step 2.
Step 2: Choose the right first use cases
The most common mistake in AI projects is to start with the most impressive use case instead of the most suitable one. Your first project must above all succeed; that is what buys you the organisation's trust for the next ones.
Three criteria for a good first use case
- Repetitive. The task is carried out daily or weekly, in a certain volume. A task that happens four times a year may be as annoying as you like, but it isn't your first AI project, because the gain can't be measured and the investment never pays for itself.
- Rule-based, or nearly so. The best candidates have a recognisable structure: an incoming email has to be classified and forwarded, a draft has to be written to a template, data has to be moved from A to B with fixed logic. AI's strength is handling the 20% variation that classic workflow automation couldn't, not inventing the process from scratch.
- Data-driven. The task builds on data you already have in digital form: emails, CRM records, documents, order history. If the input today is "what Lars has in his head", there is foundation work to be done first; that belongs in step 3.
Examples that typically meet all three criteria
- Meeting minutes and follow-up in sales: conversations and agreements are documented manually today (or not at all); it is repetitive, structured and data-driven. We've been through that whole area in our guide to automating sales processes.
- Inbound enquiries: classification, prioritisation and draft replies to customer emails.
- CRM maintenance: creating and updating contacts, deals and notes; see our walkthrough of what CRM automation is if the CRM is your starting point.
- Internal knowledge search: "where does it say something about X in our documents/contracts/quotes?"
Choose two, three at most. An AI strategy with twelve parallel initiatives is in practice a strategy about nothing.
Step 3: The data foundation: systems that talk to each other
AI software is only as good as the data it works on. This is the step where most projects are actually decided, long before anyone has chosen a model or supplier.
Systems that talk to each other
Make a simple overview of the systems your chosen use cases touch: CRM, email, calendar, accounting system, support system, document store. For each system: can it be read from and written to via an integration (API), or is it a closed island? An AI agent that has to update the CRM after a meeting needs to reach the calendar, the email and the CRM: if it can only reach two out of three, a human ends up as the glue in the middle, and then you haven't removed the bottleneck, only moved it.
This is also where you discover the shadow systems: the spreadsheet that is really the order management; the shared inbox that is really the case handling. They need to be on the map, because that is often where the real process lives.
Data quality: garbage in, garbage out, just faster
AI automation amplifies the data quality you have. If your CRM is full of duplicates, outdated contacts and half-empty fields, an AI agent will produce duplicates, outdated answers and half-empty drafts, just at a higher pace. Before you switch the power on, the most important data sources need to be cleaned up: duplicates merged, dead records archived, mandatory fields defined. We've written a whole guide to CRM data hygiene, because in practice it's the most overlooked preparatory work in AI projects.
Two more things belong to the foundation: access control (who, and which systems, may see which data?) and location (where is data physically stored, and what may leave the building?). Both become concrete decisions in step 4.
Step 4: Choose an approach: ready-made SaaS AI or an AI agent in your own instance
Once use cases and data foundation are in place, the choice is between two basic models. Both can be right: it depends on your data, your system landscape and how much control you need.
Model A: Ready-made SaaS AI
You buy a finished product with the AI built in: a support system with a built-in reply assistant, a meeting tool with built-in minutes, a CRM with built-in scoring. The advantages are obvious: quick start, no operations, and the supplier does the maintenance. The drawbacks are just as obvious: your data is processed in the supplier's environment on the supplier's terms, the AI can only do what the product can do, and each department easily ends up with its own AI tool: five systems, each with its own AI, none of which know each other.
Model B: AI agent in your own instance
Alternatively, you gather the AI layer in one place: an AI agent that runs in your own instance with your own database, and which is connected to your existing systems via integrations. The agent knows your customers, your documents and your workflows as a whole, instead of the intelligence being split across five products.
It's the model we're building ourselves: we are currently building Agent360 OS, where each customer can get their own instance; it has not been released yet. We mention it not as the answer, but as an example of what model B means: the AI works across the systems instead of inside one of them.
How to choose
Rule of thumb: if the use case is isolated to one system and contains little sensitive data, ready-made SaaS AI is often the quickest. If the use case goes across systems, or touches personal data, contracts or business-critical knowledge, that argues for an agent in your own instance, both for reasons of control and because cross-system workflows are where the gain is typically greatest. And whichever model you choose, the EU AI Act sets requirements by risk level, so check where your use case lands; we've been through the practical side in EU AI Act for Danish businesses.
Overview of the six steps in an AI strategy
| Step | Focus | Concrete output | Typical pitfall |
|---|---|---|---|
| 1. Map | Workflows and bottlenecks | Prioritised list with time estimates | Maps the org chart instead of reality |
| 2. Choose use cases | Repetitive, rule-based, data-driven tasks | 2-3 selected use cases | Picks the impressive over the suitable |
| 3. Data foundation | Integrations and data quality | Systems map + cleanup of key data | Switches the power on to dirty data |
| 4. Choose approach | SaaS AI vs. agent in your own instance | Decision per use case incl. data requirements | Five departments buy five disconnected AI tools |
| 5. Pilot → measure → scale | Limited test with measurable criteria | Basis for decision: scale, adjust or close | Eternal pilot with no success criteria |
| 6. Employees on board | Training, trust, new workflows | Clear ground rules + appointed super users | Rollout by email on a Friday afternoon |
Step 5: Pilot → measure → scale
Never roll AI out across the whole organisation at once. Choose one team, one use case, one limited period (typically four to eight weeks) and define before you start what success means.
Define the goals before the pilot starts
Good pilot goals are few and concrete: time saved per task, share of drafts that can be used without substantial rewriting, response time on enquiries, number of errors caught or made. Measure a baseline in the week before the pilot, otherwise you have nothing to compare with, and the evaluation ends in anecdotes ("it feels faster"). If your use case is in sales or customer work, link the goals to the figures you already steer by; our guide to revenue operations walks through how those metrics fit together across the funnel and systems.
Also set an error budget: how many errors of which type are acceptable? An AI that writes drafts a human approves can tolerate more errors than one that sends on its own. That distinction, draft vs. autopilot, is the most important single decision in the pilot design: always start with drafts, and only move to autopilot once the figures have shown that the quality holds.
Scale what works, close what doesn't
After the pilot period there are three legitimate outcomes: scale (the figures hold, roll out to more teams), adjust (the potential is there, but something needs to change, so run one more round with specific changes), or close (the gain fails to materialise: stop, and take the learning with you). The third outcome isn't a failure; it's the strategy working. A pilot that can never be closed isn't a pilot, it's a decision that was made in advance and just lacks its alibi.
When you scale, scale one hop at a time: from one team to two, not from one team to the whole company. Each expansion reveals new exceptions and local workflows, and they need to be caught while they're cheap to fix.
Step 6: Bring employees on board
No AI strategy survives contact with an organisation that doesn't want it. And the resistance is rarely irrational: employees have seen system rollouts before that gave more work, not less, and many have heard AI described as something that is going to replace them.
Be honest about the purpose
Say it straight, and mean it: the goal is to remove the repetitive, manual tasks so the time can be spent on what requires people: customer relationships, professional judgements, the difficult conversations. And let actions back up the words: if the hours the automation frees up are converted solely into higher targets, the organisation learns that the next automation should be resisted. If they are converted into better work, it learns the opposite.
Training: short, concrete, in their own workflow
Drop the generic AI courses. Train employees in their new workflow: here is where the draft arrives, this is how you correct it, this is how you report an error, this is where the line goes for what the AI may do. Appoint a super user in each team (a colleague, not a consultant) as the first line for questions. And make it safe to report errors: every error that's reported is free quality assurance; every error that's hushed up is a trust crisis waiting to go off.
Trust is built with transparency
Employees need to know what the AI can see, what it does on its own, and what always requires a human. Write it down in clear ground rules, one page, not a policy bible. In customer-facing functions this applies doubly: our guide to AI in customer success goes through how to keep the human in front while the AI works in the background.
When should you NOT automate?
This is the section most vendors would rather skip. But an AI strategy without exclusions is advertising, not strategy. Don't automate when:
- The process is broken. If you automate a bad process, you get a faster bad process. Fix the workflow first: it often turns out that half the steps could be removed entirely, without AI.
- The volume is too low. A task that takes three hours a month will never earn back the setup, operation and maintenance. The calculation from step 1 has to come out in automation's favour, otherwise it's technology for technology's sake.
- The cost of an error is high and irreversible. Dismissals, price approvals, legal assessments, anything that touches people's employment or health: here a human must make the decision, and the AI at most prepares it. The EU AI Act draws the same line formally; the European Commission's framework is risk-based, and applications in areas such as employment and essential services are treated as high-risk, with stricter requirements for, among other things, human oversight, while certain practices are banned outright.
- The data isn't ready. If step 3 revealed that key data is unstructured, outdated or scattered across shadow systems, then the cleanup is the project, not the automation. Skipping it just makes the errors arrive faster.
- The task IS the relationship. Some tasks look like a waste of time on a process diagram, but are really where trust is built: the follow-up call, the hand-tailored adjustment of the quote, the difficult customer conversation. If you automate them, you save minutes and lose relationships.
- You're doing it because the competitor is. Fear-driven automation without your own business case ends up as off-the-shelf licences that nobody uses. Better to wait a quarter and do it right.
An honest "not yet" in those areas isn't conservatism. It's what ensures that the areas you choose to automate get the resources and attention to succeed.
Typical mistakes in AI strategies
The same mistakes recur across industries and company sizes. If you recognise yourselves in more than two, stop and go back to step 1:
- Technology before problem. "We need to have AI" isn't a strategy. The starting point is a bottleneck, not a tool.
- Everything at once. Twelve initiatives, none finished. Choose two or three use cases and finish them before opening new ones.
- The pilot with no end date. Without defined success criteria and a fixed evaluation date, the pilot becomes a permanent state of exception that nobody dares close.
- Dirty data, full speed. Letting AI work on top of duplicates and outdated records gives great demos and wretched operation.
- Shadow AI without a framework. If the strategy doesn't give employees approved tools, they find free tools themselves, and suddenly customer data sits in services nobody has approved. Clear ground rules beat bans.
- No owner. AI initiatives without a named person responsible die when everyday life hits. One person must own each initiative, with a mandate to adjust and close.
- The gain is never harvested. Hours are freed up, but nobody decides what they are to be used for. Without that last link, the whole calculation is theoretical.
- Compliance as an afterthought. Personal data and the AI Act's requirements are thought about at the end instead of in steps 3 and 4, and then the solution has to be rebuilt just as it has gone into operation.
Frequently asked questions about AI strategy
Q: How long does it take to put together an AI strategy?
The strategy work itself (mapping, prioritising use cases and choosing an approach) can be done in two to four weeks in a mid-sized company, if the right people set aside the time. The first pilot typically adds four to eight weeks on top. What takes time is rarely the analysis but the data foundation: cleanup and integrations can easily be the biggest item in the overall process.
Q: Do we need to hire AI specialists to get started?
No, not for the first steps. Mapping workflows and choosing use cases takes business understanding, not computer science, and modern AI platforms and AI software are built to be configured, not programmed from scratch. What you need internally is a clear owner of the initiative and super users in the affected teams. The need for specialists typically arises only when you scale across many systems or build your own integrations.
Q: What is the difference between workflow automation and AI automation?
Workflow automation follows fixed rules: when a form is filled in, a case is created; when an invoice is approved, it is posted. That works well when the input is structured and the same every time. AI automation can additionally handle unstructured input: read an email, assess what it's about, write a draft reply, summarise a meeting. In practice, the best solutions combine both: the AI understands and drafts, while the rule-based workflows make sure things happen in the right order.
Q: Is our company too small for an AI strategy?
No, but the strategy has to be scaled to the size. In a company with ten employees the "strategy" might be one page: two use cases, one tool choice, one person responsible and three metrics. The principles are the same: repetitive, rule-based, data-driven tasks first; pilot before rollout; employees on board, but the process can be completed in weeks, not months. Small companies often have an advantage: shorter decision paths and fewer systems that need to talk to each other.
Q: What does it cost to get started with AI automation?
It depends on the approach from step 4. Ready-made SaaS AI typically costs a running licence per user, while an AI agent in your own instance is a platform investment that in return covers several use cases across systems. The biggest cost is rarely the software, though; it's the hours spent on data cleanup, integrations and training. That's why the order in this guide matters: a limited pilot with measurable criteria means you know the gain before you invest in full rollout.
Q: How does the EU AI Act fit with our AI strategy?
The EU AI Act is risk-based: the greater the risk an AI application poses to people's rights and safety, the stricter the requirements, from prohibited practices through high-risk applications with requirements for documentation and human oversight, among other things, to minimal-risk applications with no special requirements. For most administrative use cases (drafts, summaries, classifying enquiries) the requirements are manageable, but applications in areas such as recruitment and employment are treated as high-risk. The practical advice: assess the risk level already in step 2, when the use case is chosen, not after rollout.
Get started: your first 30 days
An AI strategy only becomes real once it has a calendar. A realistic plan for the first month: Week 1: map workflows in two or three teams and compile the prioritised list (step 1). Week 2: choose two use cases using the three criteria, and make the systems map of data and integrations (steps 2-3). Week 3: decide the approach per use case, clean up the key data the pilot will use, and define the pilot's success criteria and baseline (steps 4-5). Week 4: start the pilot in one team, with super users appointed and ground rules on one page (step 6).
It's deliberately unambitious. The ambition belongs in the scaling, not at the start. Companies that succeed with AI automation are rarely the ones with the biggest visions in the slide deck; they're the ones that chose two boring, repetitive, rule-based workflows, measured honestly, closed what didn't work, and scaled what did. That discipline is the whole strategy; the rest is tool choice.
#AI strategy #AI automation #Workflow automation #Automation of workflows #AI software
Contents
Which of your manual steps could disappear?
We map where data is moved by hand between your systems, and what can realistically be automated.
Book a mapping sessionRelated articles

Own or rent software? A decision guide for Danish businesses
Own or rent software? Cost over time, data lock-in, GDPR and exit, self-host and fair source, plus a concrete decision framework for Danish businesses.

Claude Code in a business: what an AI developer agent can do, and what it requires
What Claude Code, Anthropic's agentic developer tool, can do in a business, and what it requires in security, governance and organisation.

AI Barriers in Danish SMEs: What Is Holding Businesses Back?
In an autumn 2025 Danish SME survey, 44% had worked with AI and 10% had implemented it systematically. Five barrier themes and possible first steps.