How small companies use AI to run operations: a field guide
Public examples of small companies running operations with AI, what each one proves, and where the evidence stops.
Small companies put AI to work on operations in two shapes. Most buy a vertical product that runs one workflow end to end, from back-office admin and phone or WhatsApp reception to bookings and sales conversations, while a smaller group runs its own team of agents on systems it already owns. The public examples below cover both shapes, with the verified figure, the self-reported figure, and what each case does not establish.
How we checked
Every product description comes from the company's own page, read in November 2026. A size or outcome claim appears only when an independent outlet or a funding announcement states it, with the date next to the figure. Numbers that only the company publishes are labelled as such. Where sources disagree, both appear; neither is treated as correct. Anything we could not verify is left out.
Two kinds of example
Vertical software. The operator buys a product that runs one workflow on top of systems the business already uses. The vendor keeps the models and agent design. Adoption spreads fast because a small team can switch it on without engineering work, and the press covers the vendor's funding and client count.
An internal operating system. The company runs agents itself: its own memory, its own schedules, its own approval steps. Adoption is slower and the evidence is narrower, usually one company describing work it now ships.
A vendor's client count shows that many businesses bought the same workflow, not that it paid off inside any of them. An internal account shows what one company built, and it does not transfer.
Vertical AI software sold to operators
Haddock: back-office admin for restaurants
What it does. Haddock sells restaurant management software, and its agents automate the admin around it. Fina, the first agent, collects invoices from email, reconciles them against delivery notes and flags errors (haddock.app, November 2026).
Who it serves. Restaurants and small restaurant groups in Spain, Andorra and Portugal.
Verified figures. El Español reported on 15 November 2025 that more than 2,000 restaurants use Haddock's software, and that the team grew from 25 to almost 50 people during 2025, mainly in product, technology and AI.
Where the sources disagree. The same article says the agents can cut administrative work by up to 90%, about 15 hours a week per location, without describing how that was measured. Haddock's own page claims more than 24 hours saved a month. Fifteen hours a week works out to about 60 a month, 2.5 times the site's more-than-24 floor; neither carries a method.
What it proves. A narrow workflow, invoice reconciliation, can be sold to thousands of small businesses that would never build it.
What it does not prove. Nothing about profit. The time figures come from the company and an article repeating them; a restaurant that saves admin hours may spend them on more admin.
Heydiga: the phone and WhatsApp
What it does. Heydiga sells an AI receptionist that answers calls and WhatsApp in a natural voice, books, changes and cancels appointments, and hands over to a person when it does not know the answer (heydiga.com, November 2026).
Who it serves. Local service businesses: car dealers, salons, restaurants and clinics.
Verified figures. Emprendedores reported on 17 September 2026 that Heydiga raised a €5.5M seed round led by 6DC, with Kfund, Italian Founders Fund and Decelera, and that it had more than 200 active clients with more than 78% of conversations automated end to end. Forbes España reported a €1M round and 80 active clients in Spain on 4 December 2025.
Where the sources disagree. The site says more than 80% of what comes in is resolved without a handover. The press figure is 78% automated end to end. They measure different things, and the site's is the looser claim: cite 78% and attribute 80% to the company.
What it proves. A reception workflow is buyable by businesses with no technical team, and one product crosses several local sectors without being rebuilt for each.
What it does not prove. That the owner got the hours back in revenue. A receptionist that resolves 78% of conversations still leaves the other calls, and nothing here measures what the business did with the freed time.
Bookline: bookings across voice, WhatsApp and Instagram
What it does. Bookline runs voice, WhatsApp and Instagram agents that turn calls and messages into bookings, connected to the operator's booking engine, for restaurants, hotels, campsites and taxis (bookline.ai, November 2026).
Who it serves. Hospitality and mobility operators across European and Latin American markets (InfoHoreca, 30 September 2025).
Verified figures. InfoHoreca reported on 30 September 2025 that Bookline closed a €3.5M Series A led by ICF Capital with Sabadell Venture Capital and Faraday Venture Partners, and that it had more than 1,700 clients in its European and Latin American markets. El Periódico reported the same round the same day, adding that the tool handles around a million contacts a month.
Where the sources disagree. Bookline's own site says more than 1,200 hospitality teams have used it. The counts may cover different periods or definitions, so do not add them or pick one. No independent source states the headcount.
What it proves. Booking agents work across several channels and verticals, and investors funded the category on client counts.
What it does not prove. Conversion. A client count says nothing about how many calls became reservations, and no source measures that rate.
Kresco: WhatsApp as a sales channel
What it does. Kresco builds WhatsApp AI agents for direct-to-consumer brands and retailers: click-to-WhatsApp ads, lead qualification, sale recovery and support (kresco.digital, November 2026).
Who it serves. D2C brands and retailers that sell through messaging.
Verified figures. None. We found no independent report of Kresco's size or outcome.
What it proves. Only the product's shape: a small vendor can package sales and support into one messaging agent for brands on paid social.
What it does not prove. Anything about outcomes. The company publishes conversion and acquisition-cost claims; without an independent source or a stated method those are marketing copy, not evidence. This is kresco.digital, not kresco.ai.
Running it yourself: an internal operating system
Capchase
What it is. Capchase runs its own swarm of agents on its own infrastructure. Its agents took over work that used to arrive as requests: recurring reports, support triage, data collection. Our public case study documents it.
Verified figures. From the case study page, updated 2 November 2026: 70+ users, 10k+ tasks a week, 28 agents, 25 automated workflows, and 65% of users outside engineering. Task volume went from about 15 a week to about 7,000 in six weeks. The company has contributed 88 merged pull requests to the open-source project behind it.
What it proves. A company can run operations on agents it hosts itself, and colleagues outside engineering will use them when the work arrives in tools they already open.
What it does not prove. Capchase's engineering team is around 30 people, so this is not a proxy for a company under 100 people, or one with a tech team under about 20. The figures are the customer's own account, published with our involvement: a company report, not independent evidence.
The pattern
| Company | Workflow | Channel | Kind | Evidence level |
|---|---|---|---|---|
| Haddock | Invoice and delivery-note reconciliation, back-office admin | Restaurant management software | Vertical software | Independent client count and headcount; time saved self-reported |
| Heydiga | Inbound calls and WhatsApp, appointment booking | Phone, WhatsApp | Vertical software | Independent funding and client count; automation rate self-reported |
| Bookline | Bookings from calls and messages | Voice, WhatsApp, Instagram | Vertical software | Independent funding and client count; the site's client count is lower |
| Kresco | Sales and support conversations | Vertical software | Product only | |
| Capchase | Reports, support triage, data collection, security patches | Agents on the company's own infrastructure | Internal operating system | Company account in our own case study |
What this evidence does not show
Press figures about these companies are company-reported. No outlet audited a client count or a customer's books.
No example here has an independent return on investment, and time saved is not money: none of the sources say where the freed hours went. The strongest claim the evidence supports is that a product was bought and a count is large.
Haddock and Bookline each report figures that do not line up across sources, which reads like figures assembled for an announcement, not a measurement.
A customer logo on a vendor's site is not a case study, ours included. A logo says a company uses the product, not that it measured a result. None of the vendors above is presented here as a customer of ours.
Applying it in a small company
Start with one workflow that repeats often, has a clear input and output, and already annoys someone. Measure it for two weeks first: hours spent and error rate.
Then pick the shape. Buy a product when the workflow is standard in your sector. Run your own agents when the workflow is specific to how you work and the value sits in the judgment, not the click.
Decide where a person approves before the first run, not after the first incident. Customer commitments and anything irreversible wait for a human, and the rule is written down.
The company operating system covers the written side: who owns each workflow, what it may touch, and how it is reviewed. Our session examples show what an agent workflow looks like when it runs.
Sources
- El Español, "Restaurantes donde conviven trabajadores y agentes de IA: la revolución tecnológica llega al pequeño y mediano negocio", 15 November 2025. https://www.elespanol.com/invertia/disruptores/ecosistema-startup/startups/20251115/restaurantes-conviven-trabajadores-agentes-ia-revolucion-tecnologica-llega-pequeno-mediano-negocio/1003744008804_0.html
- Haddock, "Agentes IA para restaurantes", read November 2026. https://www.haddock.app/agentes-ia-para-restaurantes
- Emprendedores, "HeyDiga levanta 5,5 millones de euros en su última ronda de inversión", 17 September 2026. https://emprendedores.es/startups/heydiga-levanta-55-millones-de-euros-en-su-ultima-ronda-de-inversion/
- Forbes España, "HeyDiga cierra una ronda de un millón de euros liderada por KFund y los fundadores de JobandTalent", 4 December 2025. https://forbes.es/economia/840072/heydiga-cierra-una-ronda-de-un-millon-de-euros-liderada-por-kfund-y-los-fundadores-de-jobandtalent/
- Heydiga, home page, read November 2026. https://heydiga.com/
- InfoHoreca, "Bookline levanta 3,5 millones para revolucionar la IA en hostelería y expandirse internacionalmente", 30 September 2025. https://www.infohoreca.com/noticias/20250930/bookline-internacionalizacion-nuevos-productos-hospitalidad
- El Periódico, "La herramienta para gestionar reservas en hoteles y restaurantes con IA Bookline cierra una ronda de 3,5 millones", 30 September 2025. https://www.elperiodico.com/es/economia/20250930/bookline-reservas-hoteles-restaurantes-ia-ronda-3-5-millones-122125540
- Bookline, home page, read November 2026. https://bookline.ai/
- Kresco, home page, read November 2026. https://kresco.digital/
- Agent Swarm, "Capchase case study", updated 2 November 2026. https://www.agent-swarm.dev/case-studies/capchase
Related field notes
We shifted 53% of our token spending to scripts
In a month, that was $17k, and 450 agent hours saved
Technology as an Operations Moat for Lean Companies
How a lean company builds an operations moat with technology: capture repeated work in owned workflows, and keep context and skills as they accumulate.
Which Self-Hosted Agent Platform Fits a Team
A dated, sourced rule for a team buying a self-hosted agent platform: how six products differ on shared context and worker isolation.