Machinegeist
Report · free to read

Seven AI shifts, and what to do about each one this quarter.

Most SMEs now use some AI. Few use it in a way that changes how the company runs. This report collects seven shifts in the market that we also see in our own work with SMEs, with the outside evidence, our own cases, and one concrete action per shift.

Get the pdf

The whole report is on this page, below. The pdf is the same text in one file, to read offline or pass on.

AI isn’t always the answer. This pdf is being sent to you with good old software.

Or start reading below ↓

The seven shifts

Seven AI Shifts SMEs Can Act On This Quarter

Christophe Rosseel, Machinegeist · machinegeist.ai

Why this report

Most small and mid-sized companies now use some AI. Few use it in a way that changes how the company runs. The distance between those two groups is where the next few years get decided, and it is growing.

The numbers point the same way. In France, the share of SMEs using AI tripled in two years, from 13% to 40%. Across the EU, 42% of firms with up to 749 staff use AI or algorithmic management, mostly in limited and experimental ways. In Belgium, just over 28% of small companies do. The gap that should worry a Belgian SME owner is the one with the firm down the street that has started to rebuild its work around AI.

This report collects seven shifts in the market that we also see in our own engagements with SMEs. Each comes with what is changing, the outside evidence, what we see of it in our own work, and one thing you can do this month or this quarter. Our examples come from Voxdale, two other clients we do not name, and our own company.

1. You do not need to buy a company brain: own your context instead

What is changing. “Company brain” became a software category this year. For a company of five to fifty people, buying one is usually premature. The AI subscriptions you already pay for search your connected tools with each person’s own permissions. What they lack is a shared place where what the company knows is written down, owned, and readable by every agent you will deploy. That place can be a folder of plain-text files in a version-controlled repository. Building it is mostly writing.

Owning that layer also makes you independent of any single model. Models improve every few months and the best one changes hands. If your context lives inside one vendor’s product, switching costs you your memory. If it lives in your own files, you switch models with a configuration change.

The outside evidence. Nobody has won this category yet. Y Combinator put “company brain” on its list of startups it wants to fund in the summer of 2026, which is what investors do when a market is still open. Microsoft and Google are building one into their office suites, Notion and Atlassian into their workspaces, OpenAI and Anthropic from the model side, and a crowd of startups from scratch. Even a vendor in the field, Slite, found in a July survey of 149 teams that only 17% have a company brain that works; another 27% tried and gave up, almost always on upkeep rather than technology. Buying one now means betting on a winner nobody can name. Belgian surveys point at the work underneath instead. Exact’s 2026 KMO Barometer, a survey of more than 1,300 Belgian SMEs, found that half do not know well how to put AI to use, and Exact’s own conclusion is that the value lies in combining the technology with reliable data and clear business processes. In the Trends DataNews AI & Tech survey (June 2026), 30% of Belgian companies call themselves very dependent on American IT and cloud suppliers, and only 5% count avoiding lock-in as a deciding factor when they pick a technology partner. A company brain bought from one more supplier adds to that dependence.

We see it in our own work too. Voxdale’s operating model lives on GitHub as plain text: how projects run, the policies, what was decided and why, an entry per client. Every change carries an author and a reason, nothing lands without review, and the full history can be searched and reversed. It reaches every colleague as a plugin inside the AI assistant they already use. A document goes in when it will likely still be true in a few months, any colleague may read it, it is mostly text, and someone has checked it; everything else stays in the shared drive. Two other clients asked us the same question this autumn, a communication agency that already pays for a knowledge-base tool and a manufacturer that keeps everything on a shared drive. For both we wrote out the trade-offs of where their company brain should live, and left the decision with them. Our own company runs the same way as Voxdale: our decisions, case studies, sales rules and this report live as files in one repository that our agents read.

Do this month. Write down five things every new colleague has to learn the hard way: how you price, what you never promise a customer, who decides what. Put them in plain-text files in one shared, versioned place, give each a named owner, and point your AI assistant at it. Notice which questions it still cannot answer. That list is your next month’s writing.

2. Install systems, not agents or prompts

What is changing. Most AI at work still starts with one person who opens a chat, pastes in some context and copies the answer somewhere else. The next person starts from zero, and when the colleague with the good prompts leaves, the prompts leave too. Agents were supposed to fix that, but an agent on its own is a prompt that runs longer. The companies that get results install a system around the AI: a place where its data lives, work that runs on a schedule without anyone asking, rules for what an agent may do alone and what waits for a person, and a record of every outcome that the next run can use. The model inside is the part that changes fastest, and in a well-built system it is also the easiest part to swap.

The outside evidence. Agoria asked 307 leaders of Belgian technology companies about AI in September 2026. Six in ten are still experimenting, 37% are building AI into their processes, and only 6% say AI has become an integral part of how the company runs. 85% see no effect on revenue yet; the companies with more AI knowledge and more AI investment see one more often. McKinsey’s global survey of August 2026 finds the same divide: nearly three-quarters of the companies that get significant financial value from AI have fundamentally redesigned their workflows around it, against a quarter of the rest.

We see it in our own work too. Our growth engine is the clearest example we have, and it is now in use with three customers. It holds a database of the whole addressable market, scheduled jobs that bring in new data, agents that research companies and write briefs, a second agent that tries to refute each brief, a queue where a person approves or rejects every recommendation, and a log of every touch, so the next brief knows what happened the last time. People keep it running with three habits: log what happened, clear the queue each day, review the week.

One of those customers is a strategic-communication agency. Its managing partner set one condition: the system had to make the business developer’s work easier, not create new work. So the team posts what happened in the Slack channel it already used, in its own words: a meeting held, an offer sent, a contact who moved to another company. The system reads each post and records what it states as soon as every person and company in it matches one record. When it is unsure, it asks in the thread, and a question nobody answers within a week is dropped unwritten. It stores the facts, never the post. When a contact moves to another company, the system follows them there and the new employer becomes an account.

At a manufacturer of luxury products, the head of operations chased two main suppliers by hand every day. Since the end of October an agent reads his mail every morning and updates a dashboard that shows, per model, what has been ordered, what has been delivered and the deadline. A link to the company’s ordering system refreshes the orders every night. What he hears on the phone or in a meeting he adds himself, and the questions the system cannot answer wait for him on the dashboard.

Do this quarter. Pick the AI use in your company that depends most on one person’s prompts. Write down what it would need to run without that person: where the data lives, what starts it, what it may do alone, who approves the rest, and where the outcome is recorded. Build that, even small, before you add a second agent.

3. Teams learn AI fastest in go-to-market work

What is changing. Sales and marketing tolerate experiments better than finance or operations. A mistake costs less, feedback comes within days, and most teams already know their CRM data is imperfect. That is why agents show up there first, and why the skills a team builds there transfer to higher-stakes work later.

The outside evidence. Bitkom’s 2026 survey of 603 German companies with 20 or more staff found AI in sales at 14%, up from 5% a year earlier, and 11% of companies that use or plan AI already running agents. A 2026 Aalto University thesis built an AI prospecting pipeline with a Nordic software company: in a field test it reached about 90% contact accuracy and booked meetings, and the author estimates research time per lead could drop from five to fifteen minutes to under one.

We see it in our own work too. Building our own growth engine taught us more about agents, evaluation and human approval than any client project. The communication agency from shift 2 chose the same engine as its first agent build, ahead of its company brain and a customer-satisfaction agent, on an argument from its managing partner: an error in the CRM costs less than an error in invoicing. At Voxdale, a proposal now starts from a command that reads the company’s positioning, commercial policy and brand voice before drafting; it takes under an hour, where it used to take an afternoon or more, and it never sends anything without a person’s approval.

Do this month. Give one commercial person or a small team a defined pilot: lead research, account briefs, or first drafts of outreach. Agree up front what a person must approve before anything goes out (shift 7). Measure hours and quality, and present the result at the next team meeting. Leave the most regulated processes for later.

4. Tools that agents cannot talk to get replaced, because switching got cheap

What is changing. Agents work with tools that have clean interfaces, structured data and permissions they can respect. Tools that keep data locked behind a screen become the bottleneck, and the cost of leaving them has collapsed: an agent can read the old system, write the migration, and document what it did. For years the rational choice was to keep a mediocre tool because switching was a project. That calculation has flipped.

The same holds for suppliers. Taking over software someone else wrote used to be the job nobody wanted. Now that agents can read code and documentation, we take on other developers’ work without hesitation. A company that worries about being locked in to a small supplier should first ask how readable the supplier’s work is. Size matters less once someone else can pick the work up in days.

Most AI agencies and implementation partners only automate work the company already does. But most mid-market companies care more about revenue growth than time savings. Tools an agent can reach are what make the growth side possible: an agent that can read your CRM, your quotes and your project history can help find the next customer, not only file the last invoice.

The outside evidence. The 2026 Flash Eurobarometer describes AI use in European SMEs as mostly limited and experimental, which is what you get when AI sits on top of tools it cannot reach. In McKinsey’s 2026 survey, 32% of respondents say their organisation decided against buying at least one software product or feature because it could build it in-house with AI coding agents. Bitkom’s survey shows German companies using AI for internal knowledge management doubling in a year, from 11% to 22%, and that only works when the knowledge sits somewhere an agent can read.

We see it in our own work too. In January, Voxdale ran on more than forty point solutions that did not talk to each other. In the second quarter we retired Confluence: the project knowledge moved into the project drives teams already used, in eight days including design and verification, for a job that would have taken a quarter not long ago. An agent wrote and debugged the migration scripts; the scripts moved the data, the same way every time. Colleagues then built their own small tools where a subscription had been a poor fit, and one weekend project to replace an e-signature subscription ended with the discovery that the company’s existing office suite already covered the need.

Do this quarter. List your ten most-used tools. Mark which ones an agent can read from and write to through a proper interface. Pick the one that causes the most re-typing between systems and run a thirty-day test of either an agent-friendly alternative or a small integration you own. Measure the time and the errors next to the subscription cost.

5. Open models are good enough for much of the work, on your own terms

What is changing. A year ago, serious AI meant renting it from a handful of American labs. In 2026, open models that you can run on a machine in your office or on European infrastructure are able to handle a large share of everyday business work: classifying, summarising, extracting, drafting from your own documents. The frontier labs still lead on the hardest tasks, and renting remains cheaper per token. For the first time, where your AI runs is a decision you make per workload.

Fine-tuning a model on your own examples has also become cheap enough for a small team to try. We treat it as an edge case. A fine-tuned model ties your investment to one model generation, and a new generation arrives every few months. For most SMEs, well-organised company knowledge that any model can read (shift 1) does more, for longer.

The outside evidence. Running costs are part of the case. In McKinsey’s global survey of August 2026, about one in five respondents say AI running costs, token bills included, already limit how much they use AI. Tilde, a language-technology SME in the Baltics, is a case where adapting the model is the product. It used European supercomputing capacity to fine-tune models to Baltic and Slavic languages and then shipped an enterprise search and question-answering product that customers can run on their own premises or private cloud. Error rates dropped from over 27% to levels comparable with major languages, which makes the product usable in regulated sectors such as government, finance and healthcare. The heavy computing was needed only for training; the product runs on the customer’s own infrastructure. On a much smaller scale, an individual practitioner described on Hugging Face’s community blog how a small open model, trained on about 500 examples in runs of roughly fifteen minutes, went from 36.5% to 76% on a sealed exam about EU crypto regulation. Treat that one as an indication, not a benchmark: one person, one task, and the cost multiplies with every iteration.

We see it in our own work too. We run local AI on our own hardware for the data we will not send to a rented service. We also test new models against our own work before adopting them: when a fast, very cheap new model launched this autumn, we ran it on the judgments our sales engine makes every day and stayed with the model we had, because quality mattered more than speed for that use. That test took a weekend because our context and our test cases live in files we own, not inside a vendor’s tool. We published a decision guide on local and open-source AI for mid-market companies; it is free on our website: machinegeist.ai/en/local-ai.

Do this quarter. List the AI tasks in your company and mark the ones that touch data you would rather not send out. For one of them, run an open model on EU-hosted capacity or on a machine you own for thirty days, next to the service you rent today. Compare quality on your own examples, not on a leaderboard.

6. Stress-test your internet-facing infra before hostile AI agents do

What is changing. Ransomware and data leaks used to need a skilled attacker with time to spare. AI models now find and chain vulnerabilities at a speed and scale no human team matches, and an automated attacker does not pick a victim: it scans everyone and goes through whichever door was left open. For an SME, the website, the webshop, the remote-access login, the mail server, a forgotten test environment and every AI tool connected to company data all sit on that same list. Small firms rarely have a security team, and “we are too small to be a target” stops holding when the attacker is a machine checking every address.

The outside evidence. The Centre for Cybersecurity Belgium (CCB) received 588 incident reports between January and late September 2026, more than in all of 2025. In September it warned that AI models and agents find and exploit vulnerabilities far faster than human hackers; it has not yet seen an attack carried out by AI agents in Belgium, but assumes attackers already use AI to automate theirs. Its first advice is to list every system reachable from the internet, cut that exposure to the minimum, and patch actively exploited vulnerabilities first. SMEs are the weak link: in the CCB’s own market study, only 17% of micro-enterprises and 23% of small companies have a formal cybersecurity policy, and many believe they are too small to be targeted. AI tools add doors of their own. When Google’s Mandiant team tested a bank’s chatbot before launch, they did not need clever prompts: they found an exposed interface open to a classic database injection, slipped fake system messages into a chat history the app trusted, and got unauthorised account changes through. OWASP now publishes a quarterly round-up of real exploits against generative AI systems.

We see it in our own work too. We also point AI at our own attack surface, the way an attacker would. Since the end of October, a scan on our own server checks every night what we run (software packages, automated workflows, background jobs, open ports, stray secrets) against public vulnerability feeds; only package names leave the building. The next layer is in build. A local model on the same machine reads each new vulnerability and scores it against what it knows about our architecture and the systems we run for clients. Only the items it marks as “affects us” go to a frontier-model agent, which works out the fix, and a person approves the change. The scoring stays local on purpose: a list of your own weak spots is the last thing to send to a rented API. In a test on our own hardware, three local models each caught all seven cases that needed action the same day and invented no vulnerability ID in 280 answers; fixed rules alone caught four.

Do this month. List everything your company exposes to the internet: website, webshop, remote access, mail, cloud apps, test environments, and every AI tool connected to company data. Ask your IT partner which of them run software with known, actively exploited vulnerabilities, and whether multi-factor authentication is on everywhere. Then spend one hour with whoever built your AI tools trying to make them leak data or ignore their instructions. Fix the three worst findings and put the next round in the calendar for next quarter.

7. Human oversight moves from rubber stamp to real control

What is changing. The first wave of AI at work followed one pattern: the AI drafts, a person glances, the person approves. That pattern gives a feeling of control without the substance. A reviewer who sees a hundred drafts a day, does not know how the model fails, and has no authority to stop the process adds a signature and little else.

The useful question is where a person sits relative to each process:

Position What the person does Fits Example
In the loop Reviews every output before it goes out Anything customer-facing that carries your name, until you trust it A quote, a contract clause or a reply to a complaint: the agent drafts, the account manager reads every one and sends it
On the loop Watches the flow and steps in on exceptions Internal work with clear rules and visible errors Matching supplier invoices to purchase orders: the agent books the clean matches, the bookkeeper handles the mismatches it flags
Near the loop Spot-checks samples and owns the escalation High-volume work where an error is cheap to fix Sorting incoming support mail or scoring new security alerts overnight: a person reads a sample each week and decides what gets escalated

The outside evidence. In April 2026, Sullivan & Cromwell, one of Wall Street’s best-known law firms, apologised to a bankruptcy judge for a motion with about three dozen errors, among them case citations its AI tool had invented. The firm had everything on paper: mandatory AI training, a rule that every AI answer must be checked, a review step for citations. Its partner wrote to the judge that the review “did not identify the inaccurate citations generated by AI”. The other side’s lawyers found them. Replace the judge with a client and the motion with a proposal, and any agency owner knows how that week goes. Too much review fails as well. In G2’s survey of 1,019 software buyers (October 2026), 31% say staff spend six or more hours a week supervising agents, and 45% say oversight and rework eat more than a quarter of the gains they expected. Rubber-stamp review lets the errors reach the client; reviewing everything by hand means the agent stops paying for itself.

The EU AI Act makes meaningful oversight a legal duty only for high-risk systems (Articles 14 and 26), but the Spanish AI supervisor’s guide on human oversight is the most practical description we know of what meaningful looks like, and it reads well for any company.

We see it in our own work too. Voxdale, a client of ours, designs and engineers physical products. There, everyone can read the company’s operating model through their AI assistant, but nobody can change it on their own: a proposal arrives, a named owner per domain approves it, and only then does it count. The assistant reads with each person’s own permissions, so it never widens anyone’s access, and it writes nothing back into the shared drive. Our sales engine works in two tiers. A short list of top accounts (Tier A) gets account-based marketing: agents research each company in depth and we approach it one by one. The wider market (Tier B) gets a volume approach with lighter research per company. In both tiers, agents research companies and draft proposals all day, and nothing goes out until a named person approves it. At the manufacturer from shift 2, the agent that reads supplier mail flags every change in payment details, such as new bank details on an invoice or a supplier account said to be closed, for the head of operations to check. We designed each of these gates before we built the agents behind them.

Do this month. Pick one workflow where AI output reaches a customer or a decision. Write on one page who sees the output before it goes out, under which conditions, and who has the authority to stop it. Run it for two weeks, then move the person one position out if the errors allow it.

A 90-day sequence

When What Shift
Week 1 Map where AI is already used; write the oversight rule for every workflow where AI output reaches a customer 7
Week 2 Map everything reachable from the internet, patch what is actively exploited, and stress-test every AI tool that touches customers or company data 6
Weeks 3-4 A go-to-market pilot live, with an approval gate 3
Month 2 The core of the company written down (how you sell, deliver and decide), owned and readable by every agent 1
Month 2 Replace or connect the tool that causes the most re-typing 4
Month 3 One system in production, with its own data, a schedule, an approval gate and a named owner 2
Month 3 First workload on an open model 5

The SMEs we see pulling ahead rarely have the largest AI budgets. They treat these shifts as decisions about how the company runs, made by the people who run it, and they start with the workflow that already hurts. If you want to talk through where your company stands, book a call on machinegeist.ai.

Sources

# Source Used for
1 France Num barometer 2026, September 2026 (n=9,655). francenum.gouv.fr 40% / 13%
2 European Commission, Flash Eurobarometer on AI in companies, 14 Oct 2026. ec.europa.eu 42%, “mostly limited and experimental”
3 Statbel, AI use by Belgian companies (2025 data). statbel.fgov.be Belgium 34.5%, small firms just over 28%
4 Y Combinator, Requests for Startups, Summer 2026. ycombinator.com/rfs Shift 1
5 Slite, Company brain solutions in 2026, September 2026 (survey of 149 teams, July 2026; vendor source). slite.com Shift 1
6 Exact, KMO Barometer 2026 (n=1,333 Belgian SMEs, Q1 2026). exact.com Shift 1
7 Trends DataNews, AI & Tech survey on cloud use, 4 June 2026. trends.knack.be Shift 1
8 Machinegeist, Building Voxdale’s company brain, Sept 2026. machinegeist.ai Shifts 1, 3, 4, 7
9 Agoria, Make AI Work, 30 September 2026 (n=307 Belgian technology companies). agoria.be Shift 2
10 McKinsey, The state of AI in 2026: On the road to ROI, 25 August 2026. mckinsey.com Shifts 2, 4, 5
11 Bitkom, Erstmals nutzt Mehrheit der Unternehmen KI, 14 September 2026 (n=603). bitkom.org Shifts 3, 4
12 Tuomisto, Enhancing B2B prospecting with AI, Aalto University, April 2026. aaltodoc.aalto.fi Shift 3
13 FFplus success story: Tilde’s locally deployable enterprise search. ffplus-project.eu Shift 5
14 Small LLM as a MiCA specialist, Hugging Face blog (individual author). huggingface.co Shift 5
15 Machinegeist, Local and open-source AI decision guide. machinegeist.ai/en/local-ai/ Shift 5
16 CCB, Figures show a steady rise in incidents, 15 October 2026. ccb.belgium.be Shift 6
17 CCB Safeonweb, AI accelerates and amplifies cyber threats, 29 September 2026. atwork.safeonweb.be Shift 6
18 CCB / NCC Belgium, Strategic market research into the Belgian cybersecurity sector, 2026. ccb.belgium.be Shift 6
19 Google Cloud / Mandiant, red teaming AI applications, June 2026. cloud.google.com Shift 6
20 OWASP GenAI, Exploit Round-up Report Q1 2026. genai.owasp.org Shift 6
21 Reuters, Sullivan & Cromwell apologizes for AI errors, 21 April 2026. reuters.com Shift 7
22 G2, 2026 AI Agents Report, 14 October 2026 (n=1,019). prnewswire.com Shift 7
23 AESIA (Spanish AI supervisor), Guide 6: Human oversight, July 2026. aesia.digital.gob.es Shift 7
24 EU AI Act, Article 14, AI Act Service Desk. ai-act-service-desk.ec.europa.eu Shift 7

Get the pdf

The whole report is on this page, below. The pdf is the same text in one file, to read offline or pass on.

AI isn’t always the answer. This pdf is being sent to you with good old software.

Start a conversation

How can we help?

Ghent office

Lindestuk 15

9032 Ghent

Antwerp office

De Wittestraat 60

2600 Berchem