Every January, the automation industry publishes predictions, and every January most of them describe the vendors' roadmaps rather than the customers' reality. This survey is different in intent, at least. It is built from what we at BTAC actually saw across SMB client engagements through 2025 — the workflows we built, the ones we told clients not to build, the platforms whose pricing changed mid-project, and the maintenance tickets that told us what breaks in production. Some of what follows is genuinely new. Some is old ground that the hype cycle has finally caught up with. We will try to be clear about which is which.
Seven trends, each with the same structure: what it is, what we see in real client work, and what to do about it. Two of the seven are about what is stalling, because knowing what not to buy is worth more this year than knowing what to buy.
Trend 1: AI steps move inside mundane workflows
What it is. The durable change in 2025 was not autonomous agents. It was something less photogenic: large language models embedded as individual steps inside ordinary workflows. A Make.com scenario that extracts line items from a PDF invoice into structured fields. An n8n workflow that classifies inbound emails into "quote request", "complaint", or "noise" and routes them. A Zapier step that drafts a first-pass reply for a human to approve. Extraction, classification, drafting — three verbs that describe most of what AI reliably does in production today.
The pattern that works is consistent: AI handles the unstructured seam between two structured systems. Documents in, fields out. Messy text in, a category out. A human stays on one or both ends of the step, and the workflow has a deterministic fallback when the model's confidence is low.
What we see in client work. In one engagement with a commercial cleaning company, inbound quote requests arrived as free-text emails in every imaginable format — square footage buried in the third paragraph, site address in a signature, sometimes just a phone number. The old process was a coordinator reading each email and typing a record into the CRM, about ninety minutes a day. We inserted an extraction step: the email body goes to a model with a strict field schema, the result lands in Airtable as a draft record, and the coordinator reviews a queue instead of a keyboard. Review time is about twenty minutes a day. The interesting detail is the error mode: the model is wrong in a different way than a tired human is — it confidently invents a square footage rather than leaving it blank — so the workflow flags low-confidence fields for mandatory review rather than trusting the output.
That detail generalizes. Every successful AI step we shipped in 2025 had three components: a narrow, well-specified task; structured output validated before anything downstream consumes it; and a human review point positioned where the cost of a wrong answer is high. Every failed pilot we were called in to fix was missing at least one of the three.
One practice worth stealing: measure the step like you would measure a new employee. Before an AI step goes live, we run it against fifty to a hundred historical examples where the correct answer is known, and we record the accuracy by field, not in aggregate — a model that extracts the email address perfectly but mangles the date in eleven percent of cases is a model whose date output needs a format check or a mandatory review flag. Aggregate accuracy hides exactly the field that will hurt you. After go-live, we keep sampling the outputs weekly for the first month, because the failure profile on live data is never quite the same as on the test set.
What to do about it. Inventory the places where unstructured data enters your operation — emails, PDFs, photos of forms, voicemail transcripts — and where a human currently converts it into typed fields. Those seams are the highest-value AI insertion points. Start with extraction, because it is verifiable: the answer is either in the document or it is not, so you can measure accuracy in a week. Defer drafting tasks that go directly to customers; keep those human-approved. And budget the review step honestly — an AI step that saves forty minutes but creates thirty minutes of anxious double-checking has saved ten minutes, which may still be worth building, but only if you knew that going in.
Trend 2: Platform consolidation and pricing pressure
What it is. The automation platform market spent 2025 doing what maturing software markets do: consolidating and repricing. The major platforms continued moving their pricing away from simple task counts toward models that charge more precisely for what heavy usage actually costs them — and "more precisely" means heavier users pay more. AI steps, which consume real compute, accelerated this: a workflow with a model call in the middle is no longer a few cents of tasks, and vendors price accordingly. Meanwhile, the big suites — the CRMs, the accounting packages, the commerce platforms — keep shipping native automation features that absorb the simplest use cases their customers were paying a third-party tool to perform.
What we see in client work. Two patterns. First, cost reviews that did not happen. A client built forty Zapier zaps over three years at modest volume, then added a customer portal that multiplied their task volume by nine in one quarter, and the invoice tripled before anyone noticed — the finance lead found it, not the operations lead, which tells you how invisible platform spend is until it is not. We moved the six highest-volume workflows to n8n on a small virtual server and cut the recurring bill by more than half. Second, redundant automation. We now routinely find clients paying for a platform scenario that replicates a feature their CRM shipped last year — native quote follow-up sequences, for instance — because nobody re-audited the stack after the vendor update.
What to do about it. Once a year, put your automation platform invoices next to your task volumes and compute your effective cost per thousand operations. If the number has drifted up, you have three levers: consolidate redundant scenarios, move high-volume workflows to a platform with friendlier pricing at scale, or downgrade features that a suite now gives you natively. Also, treat platform choice as per-workflow, not per-company. The right answer in 2026 is frequently a mix — Zapier where a niche connector saves a week of custom work, Make.com where branching logic is dense, n8n where volume or data sensitivity justifies self-hosting. Insisting on one platform for everything is a purchasing habit, not an architecture.
Trend 3: Self-hosted and private automation gains ground
What it is. n8n's continued rise among SMBs is the clearest signal here, but the driver is not cost alone — it is data control. As workflows started carrying customer records, financial data, and documents through AI steps, a question that used to be reserved for enterprises arrived in twenty-person companies: which third-party servers does our data pass through, and under what terms? Self-hosted automation — n8n on your own server or a private cloud instance — is the SMB-scale answer. The workflow engine runs on infrastructure you control, credentials live in your vault, and calls to external AI APIs become an explicit, auditable choice rather than an ambient default.
What we see in client work. The trigger is usually a specific event rather than a policy. A medical-adjacent services client discovered that a well-meaning employee had built a Zapier scenario passing intake-form contents — including health details — through an AI summarization step. Nothing bad happened, but the compliance review that followed ended all cloud automation experiments for six months, which was its own cost. We rebuilt the legitimate workflows on self-hosted n8n: intake data stays on their server, and only de-identified fields ever touch an external model. Elsewhere, the driver is simpler: Baserow and n8n on the same small server gives a business a database and an automation engine with no per-row or per-task ceiling, and for a client processing tens of thousands of order events a month, the flat hosting cost beats any per-operation pricing by a wide margin.
The honest counterpoint: self-hosting is not free, it is differently priced. Someone must own updates, backups, monitoring, and the 2 a.m. question when the server disk fills. We see DIY self-hosting fail when that ownership is unassigned — the workflow works for eight months and then silently stops, and nobody knows for three weeks.
What to do about it. Self-host when one of three conditions holds: you handle regulated or sensitive data, your operation volumes make per-task pricing painful, or you already have someone competent to own a server. If none hold, stay on hosted platforms and instead do the cheaper work of data hygiene — know which fields flow through which vendor, and strip sensitive fields before AI steps. If you do self-host, buy the maintenance explicitly, whether from an internal person with the time protected or from a service like ours. An unowned self-hosted stack is worse than a hosted one, because hosted platforms at least fail loudly.
Trend 4: Operational databases become the backbone
What it is. The least glamorous trend on this list and, in our view, the most consequential. SMB operations are consolidating around a small number of structured databases — Airtable and Baserow chief among them — that act as the operational backbone: the place where jobs, orders, clients, and statuses actually live. Automations then read from and write to that backbone instead of passing data hand-to-hand between apps. The shift sounds technical but the effect is organizational: when there is one table of record for "jobs in progress", the "just checking on this" messages stop, because the answer is a filtered view, not a person.
What we see in client work. The before picture is nearly identical across clients: the truth about any given order is distributed across an inbox, a spreadsheet, someone's memory, and the accounting system, and reconstructing it takes a message to two people. The after picture, when we build it right, is an Airtable base where every order is a record with a status, every status change fires a notification through Make.com or n8n, and every report is a view rather than a weekly assembly exercise. In one engagement with an events company, moving job tracking from a shared spreadsheet to a structured base eliminated an entire category of error — two crews dispatched to the same site — because assignment became a field with a constraint rather than a colored cell with a convention. Choosing between Airtable and Baserow has become a real decision: Airtable for polish, interfaces, and ecosystem; Baserow when row limits, self-hosting, or data residency matter. We deploy both and the decision is usually made by the pricing table at the client's projected volume.
A second development inside this trend: the interface layers on these databases got good enough in 2025 to retire a category of small internal tools. Airtable interfaces and Baserow's application features now cover the common cases — a dispatcher's daily queue, a form for site staff on phones, a read-only client view — that previously justified a separate Softr or Webflow build or even a small custom app. We still deploy Softr where a polished client portal matters, but the threshold for "we need a custom tool" moved noticeably upward. Fewer moving parts, fewer vendors, fewer failure points.
What to do about it. Before your next automation, ask where its data will live. If the answer is "in the workflow tool" or "in a spreadsheet someone emails around", stop and put a real database underneath first — automation without a system of record amplifies chaos rather than removing it. Design the base around statuses and handoffs, not around how the old spreadsheet looked. And give the team the views before you give them the automations; people trust a system that answers their questions before they will trust one that acts on their behalf.
Trend 5: Automation maintenance becomes a service
What it is. A quiet professionalization. As SMB automations moved from "nice experiment" to "the invoicing stops if this breaks", the question of ownership became unavoidable. Automation maintenance as a service — monitoring, error handling, connector updates, and a named person responsible when a workflow fails at month-end — emerged as a real category in 2025. It is unglamorous work: watching failure logs, re-authenticating expired API tokens, adjusting a scenario when a vendor changes a field name, and keeping a runbook so the knowledge does not live in one person's head.
What we see in client work. The failure mode is so consistent we can schedule it. A business builds automations, the enthusiastic internal builder leaves or gets busy, and over the following year the workflows decay: an API token expires and order syncs stop; a Shopify update renames a field and the Xero mapping silently drops a line; a new product variant bypasses the rules entirely. Nobody notices until a customer or an accountant does. Across our maintenance engagements, the single most common first ticket is "we think this has been broken for a while." The fix is rarely technical — it is the absence of monitoring. A workflow that fails loudly, with an alert to a named owner, is a solved problem; the same workflow failing silently is a liability accumulating interest. We now treat alerting and a monthly review as part of the build, not an optional extra, and roughly a third of our ongoing revenue is keeping automations we — or someone else — built running correctly.
What does a maintenance arrangement actually look like in practice? Less than people imagine, which is the point. Ours run as a monthly cycle: automated monitors on every workflow (error alerts routed to us, not to the client's inbox), a one-hour review of failure logs and volume drift, a short written summary of what broke and what changed, and a standing window for small adjustments — a new product line added to the routing rules, a field renamed after a vendor update. The client-visible output is a page a month and the absence of surprises. The invisible part is the runbook we maintain for every workflow, so that the knowledge of how the operation's plumbing works survives staff changes on both sides.
What to do about it. For every workflow in production, write down three things: what it does in one sentence, who owns it, and how you would know if it stopped. If any answer is blank, that workflow is unowned and will fail silently eventually — add failure notifications at minimum, which most platforms make easy and most builders skip. If you have more than a handful of business-critical workflows and no internal person with protected time to own them, buy the maintenance. It is far cheaper than the month-end incident where invoices did not go out for nine days.
Trend 6: What is stalling — autonomous agents and rip-and-replace ERP
What it is. Two hyped categories that, in our SMB client work, mostly did not deliver in 2025. First, fully autonomous agents: systems given a goal, a browser, and access to your tools, left to complete multi-step work unsupervised. Second, rip-and-replace ERP projects: the decision to fix operational chaos by buying one large system that does everything, replacing the tangle of apps and spreadsheets in a single program.
What we see in client work. On agents: we tested them seriously, because clients asked. The honest summary is that an agent completing a ten-step browser task succeeds often enough to demo beautifully and fails often enough to be unusable without supervision — and once you add the supervision, you have rebuilt the human-in-the-loop workflow from Trend 1 with extra steps. The failure modes are not exotic: a supplier portal changes its layout, a login expires, the agent misreads a total, and the cost of checking its work approaches the cost of doing it. We know of zero SMB production deployments, ours or others', where a fully autonomous agent runs a business-critical process unsupervised. On rip-and-replace: we watched two clients attempt it in adjacent industries. Both projects ran roughly twice the planned time; one went live with data migration errors that took a quarter to untangle; both discovered, eighteen months in, that the new system still did not handle the two or three processes that made their business peculiar — so they ended up with the ERP plus a tangle of workarounds, which is the old architecture at ten times the price. The mid-sized businesses that succeeded with big systems all did it the slow way: one module, in production, stable, then the next.
What to do about it. On agents, adopt the useful half of the idea — AI doing bounded sub-tasks with validation and human checkpoints — and decline the unsupervised half for anything touching money, customers, or inventory. Revisit in twelve months; the capability is improving, and the right posture is scheduled re-evaluation rather than permanent dismissal. On ERP, if your operational pain is real, the evidence from our client work is that staged automation on your existing stack recovers most of the hours at a fraction of the cost and risk. Fix the three processes that cost the most, run them for six months, and only then ask whether the remaining pain justifies a platform migration. Sometimes it does. More often, the migration urge fades once the worst manual work is gone — which suggests the urge was about the manual work, not the platform.
Trend 7: Where the next twelve months are heading
What it is. Extending the trends above without inventing new ones. We expect four developments to matter for SMB operations through 2026: AI steps becoming a standard, priced feature of every major automation platform rather than a differentiator; continued pricing pressure pushing more high-volume and data-sensitive workloads toward self-hosted stacks; operational databases absorbing more of what used to require a custom internal tool, as their interface layers improve; and maintenance — monitoring, ownership, runbooks — becoming the normal way serious businesses buy automation, the way they already buy bookkeeping.
What we see in client work. The leading indicators are already visible. Clients now ask about data flows before asking about features, which did not happen two years ago. The briefs we receive increasingly say "we have automations, they are decaying, we need ownership" rather than "we want to start automating." And the conversation about AI has matured from "can it do our quoting" to "which fields can it extract and who checks them" — a much better question, and one with a buildable answer.
What to do about it. Position for the boring version of the future, because the boring version is the one that arrives. Build your system of record before your automations. Insert AI steps at the unstructured seams with validation and human review. Choose platforms per workflow, and re-price your stack annually. Name an owner for every workflow in production, or buy the ownership. None of this requires predicting which vendor wins; all of it pays back under every plausible version of 2027.
Where to start
If this survey has a single message, it is that the gap between businesses that benefit from automation and those that merely buy it has nothing to do with access to technology — the tools are cheap and available to everyone. The gap is sequencing and ownership: measuring the manual work first, building the highest-value fix rather than the flashiest one, and assigning someone to keep it running. A two-week audit of where your team's hours actually go will tell you more than any trend report, including this one.
If you want that picture for your own operation, book a 30-minute audit. Bring your platform invoices and your list of suspected time sinks — we will tell you which of these trends apply to you, which you can safely ignore, and what the manual work is costing you while you decide.