Agents that take real actions need limits, not enthusiasm. The seven controls worth building, why each exists, and which ones you can skip.
·4 min read
There is a meaningful difference between an assistant that answers questions and
an agent that takes actions. The first can be wrong and waste a minute. The
second can be wrong and email four hundred customers.
Most of the writing about agents is about capability. This is about the other
half, which is what stops a capable thing from doing something expensive.
Start from the blast radius
For every action an agent can take, answer one question: if this fires
incorrectly a hundred times before anyone notices, what does that cost?
The answers sort actions into three groups, and each group gets different
treatment.
Free to be wrong. Reading a record, drafting text nobody sends, running a
report. Let the agent do these without ceremony.
Recoverable. Creating an internal ticket, updating a status, adding a note.
Allow, log, make reversible.
Not recoverable. Anything that reaches a customer, moves money, deletes
data, or changes access. A human confirms, every time, with no exception for
convenience.
The mistake is treating all three the same, in either direction. Requiring
approval for reads makes the agent useless. Not requiring it for refunds makes
it dangerous.
The seven controls worth building
1. A confirmation step that shows the actual payload. Not “send email to
customer?” but the recipient, the subject and the body. Approving a description
of an action is not approving the action, and people click yes on vague prompts.
2. Idempotency. Agents retry. Networks fail mid-call and the agent tries
again, which is correct behaviour and produces two invoices unless every action
carries a key that makes the second attempt a no-op.
3. Rate limits on the agent itself. Not on users, on the agent. A loop that
should have run once and ran two hundred times is the single most common
expensive failure, and a hard ceiling per hour turns it into a small bill and an
alert.
4. A hard stop. One switch that disables all actions and leaves reads
working. Somebody who is not the developer needs to be able to flip it at 7pm.
5. Full logging, in business terms. What was asked, what was decided, what
was called, what came back, who approved. Written so a manager can read it, not
only a developer. This is what you use when a customer asks why they received
something.
6. Separation of read and write credentials. The read path and the write path
should not share an account. It costs nothing and it means a mistake in the
read-only half cannot write anything.
7. Untrusted input handled as untrusted. If the agent reads a customer
message, a web page or a document, that content can contain instructions. Treat
anything the agent reads as data, never as direction, and never let content from
a record widen what the agent is permitted to do. Narrow tools are the practical
defence here, which is one reason
MCP servers should expose
little.
What you can skip
Not everything marketed as agent safety earns its place in a small deployment.
Elaborate permission matrices. Two or three roles cover most small
businesses. A configurable policy engine is a system to maintain in exchange for
flexibility nobody uses.
Model-based filtering of the model’s own output. Useful at scale, mostly
theatre for a team of twelve. A human approving the three risky action types is
cheaper and more reliable.
Real-time dashboards. A daily summary email of what the agent did is read.
A dashboard is opened twice and then never again.
Full autonomy as a goal. An agent that drafts and a human who approves is not
a failed agent. For most small business workflows it is the correct end state,
because the approval takes four seconds and removes the entire category of
expensive mistakes.
The failure that gets missed
The dramatic failures get attention. The quiet one is worse: an agent that
works, is trusted, and slowly drifts as the underlying system changes around it.
Field meanings shift, a status value gets reused, and the agent keeps acting
confidently on an assumption that stopped being true.
The defence is sampling. Once a week, somebody reads ten of the agent’s actions
end to end and confirms they were right. It takes fifteen minutes and it is the
only control on this list that catches drift, because nothing about drift throws
an error.
Building it in the right order
Read-only first, for a fortnight. Then one recoverable write, logged. Then one
approval-gated action. Each step stays in place long enough to be boring before
the next one is added.
Teams that go straight to autonomous action either get lucky or get a story
they tell later. There is no third outcome worth planning for.
If you are scoping something that will take actions rather than just answer
questions, describe what it would be allowed to do. The interesting
part of that conversation is always the list of things it should not.
One email when something new goes up, roughly twice a month. Plain
writing on what software costs and what is worth building. No sequences,
no sales calls, and one click to leave.
Tell us what the problem is. You will get an honest read on whether it is worth building, what it would take, and roughly what it would cost. No pitch deck, no pressure.
Six things: marketing websites, web applications, internal tools, AI systems, AI agents and MCP servers, and database work. Most projects touch two or three of them rather than sitting neatly in one.
Yes. Marketing sites, landing pages and content platforms, built on modern static tooling rather than page builders. They load in under a second and have the technical SEO groundwork done properly.
Yes. Customer portals, dashboards, booking systems and SaaS products. The kind where users log in, data persists and money changes hands, so it has to be right.
Software your staff use rather than your customers. A dashboard pulling numbers from systems that do not talk to each other, a job tracker matching how you actually work, or a scheduled job replacing a report someone assembles by hand every Monday.
Yes, and we will tell you when it is the wrong tool. Retrieval over your own documents, support assistants, document extraction and classification, all built with evaluation and guardrails from the start.
MCP is the standard that lets AI assistants reach your real systems instead of guessing. We build MCP servers that expose your data and tools to Claude and other clients, and agents that carry out multi-step work with permissions and audit trails.
Yes. Schema design, query performance work, migrations between systems, ETL pipelines and reporting layers. Usually called in when the database has quietly become the bottleneck, or the risk.
Not native iOS or Android apps. We build web applications that work properly on phones, which covers most of what businesses actually need. If you need a true native app, we will say so rather than talk you into a compromise.
Related
We design and build the sites and applications we make. We do not do standalone brand identity, logo design or print work. If you already have brand guidelines we will follow them.
Related
Often, yes. The first step is a short paid audit of the existing code and infrastructure, so you get an honest assessment of what is salvageable and what it would cost either way before committing to anything.
Related
Yes, if you want it. You can also take the project and run it yourself using the documentation provided. Both are genuine options and there is no pressure toward a retainer you do not need.
Related
A marketing site typically lands between $3,000 and $12,000 depending on page count and complexity. You get a written fixed price before any work begins, so the number you approve is the number you pay.
Applications and internal tools are scoped individually and most start around $15,000. Once users log in and data persists, you are buying software rather than a website, and it is priced like software.
Related
A single automation or integration is usually $2,000 to $6,000. A proper internal application with its own database and user roles runs $6,000 to $20,000. Anything your operations depend on goes up from there.
Fixed price wherever the scope can be fixed. That means we absorb the overrun rather than you, and it is usually worth the small premium on a first project together. If something changes mid-build you hear about it before the work happens, not on the invoice.
Related
A static marketing site is commonly free to host and stays free at traffic levels most local businesses never reach. A well-built internal tool usually runs $10 to $50 a month. Budget 15 to 20 percent of the build cost annually for maintenance.
Related
Copywriting, photography, ongoing SEO work and third-party subscriptions are the four things most commonly assumed by the client and excluded by the quote. Ours states what is and is not covered explicitly.
Related
Yes, typically a portion up front with the remainder on delivery, or split across milestones for larger projects. The exact split is agreed in the written proposal before work starts.
Related
Rarely page count. Integration complexity, how much the process changes mid-build, and how clean your starting data is. Messy historical data is the single most common reason projects overrun.
Related
Usually. Tell us the number and we will tell you honestly what fits inside it and what does not. That is a more useful conversation than pretending everything fits.
Related
Yes. Scoping conversations and written proposals cost nothing and carry no obligation. We will also provide a written quote for a grant application without charging for it.
Related
Yes, and they are meaningful. Registered charities, non-profits and small local businesses get reduced rates on every service, because a tight budget should not be the reason an organisation is stuck with software that does not work.
Registered charities and the volunteer-run groups that support them, non-profits and community organisations, and small owner-operated businesses including new ventures still finding their feet.
Related
There is no application and no paperwork. Mention it in your first message and the reduced rate is built into the quote you receive.
It depends on the project and the organisation, and you are told exactly what it is before committing to anything. We do not publish a single percentage because it would be either misleading or useless.
Related
Often, yes, and they are widely underused. Google for Nonprofits, TechSoup Canada, community foundations and your existing funders are all worth approaching. We wrote a guide to what is actually available.
If you are small and owner-operated, yes. Ask when you get in touch and we will tell you straight rather than making you guess.
Related
Four stages. Scope: a conversation about the problem, not the tech. Plan: a written proposal with price and timeline. Build: short cycles with something working at the end of each one. Ship and support: deploy, hand over documented code you own outright.
A straightforward marketing site takes two to four weeks from kickoff. Applications and internal tools usually run six to twelve weeks, delivered in cycles so you see something working every week rather than waiting until the end.
Related
Completely. Code lives in your repositories, hosting and domains are registered in your name, and everything is documented at handover. There is no licence to keep paying and no lock-in forcing you to stay.
Related
You keep everything, because it was always in your accounts. Documentation and a clean handover are part of the work rather than an extra, precisely so that another developer can pick it up.
Related
Most of it is at the start, agreeing what gets built. After that, expect a short check-in each cycle to look at progress. The most common cause of delay is waiting on content from the client side, so having copy and images ready helps more than anything else.
Related
Yes, and it is normal. You hear the cost and timeline effect before the work happens, so a change is a decision you make rather than a surprise on the invoice.
Related
Yes. Content is set up so you can change text and images without a developer in the loop, using either a CMS or markdown depending on what suits your team.
Related
Yes, routinely. Send yours over or ask and we will provide one.
Related
Often. Being brought in for a specific piece, or to work alongside an in-house team, is common and works fine as long as the boundaries are clear up front.
Related
A description of the problem. Not a specification, not a wireframe, just what is going wrong or what you are trying to do. Everything else comes out of the scoping conversation.
Astro and Next.js for sites and applications, React and TypeScript on the front end, Node.js and Python on the back, PostgreSQL and MongoDB for data, and Netlify or Vercel for hosting. Chosen per project rather than by habit.
Related
Only when it is genuinely the right answer, which is less often than it is used. For most marketing sites a static build is faster, safer and cheaper to run. For a large store or complex memberships, WordPress earns its place.
Wherever you need it. AWS and Google Cloud both have Canadian regions in Montreal and Toronto, and hosting in Canada is straightforward and usually costs nothing extra. If your customers ask where their data lives, that is worth having a clear answer to.
Related
Usually not. PIPEDA does not require personal data to be stored in Canada, provided you are transparent and the data receives comparable protection. Rules that do mandate it are mostly provincial public-sector rules. Health data and government contracts are stricter, so check those specifically.
Yes. Sites are built to WCAG 2.1 AA, which covers keyboard navigation, screen readers, colour contrast and focus states. It is part of the build rather than a paid extra.
Related
Sub-second loads and 95 or better on Lighthouse are the target, and normally the result. Static output with no render-blocking JavaScript is most of how that is achieved.
Related
The technical groundwork, yes: structured data, sitemaps, clean semantics, fast pages and unique metadata per page. Ongoing content and link building is a different discipline and we will say so rather than sell it.
Related
A static site has no database to breach, no admin login to brute force and no plugins running server-side code, which removes most of the attack surface. Where an application does need those, authentication and permissions are built to hold up.
Related
Usually. If your CRM, accounting package or booking system has an API or webhooks, it can be connected. Integration between systems you already pay for is one of the highest-value things to automate.
Related
Backups you have actually tested, not just configured. Restoring from a backup nobody has ever tried is a plan with an unknown failure rate.
Related
Often not, and we will say so. It is worth it when the task involves unstructured language, the rules are fuzzy, you can tell right from wrong, and a wrong answer is survivable. Fail any of those and something simpler is the better build.
Answering questions over your own documents, extracting structure from messy input like invoices and forms, triaging and routing incoming work, and generating first drafts a human then edits. Those four cover most of the real value.
Related
Chatbots on the front page, which almost nobody wants. Anything needing exact numbers, because models are not calculators. And automating a process nobody has written down, which just scales the confusion.
Related
Less than most people expect. A well-built retrieval system for a small business commonly runs in the tens of dollars a month. Costs balloon from stuffing whole documents into every request and using the largest model for tasks a smaller one handles.
Related
It can, which is why systems are built to answer from your data with sources attached rather than from memory. Anything that matters gets a human in the loop, and evaluation is built in from the start so you know when it works.
Related
A system that takes multiple steps toward a goal without being prompted at each one. Impressive in demos, frequently disappointing in production, and the gap is engineering rather than model quality.
Worth it when your team already uses an AI assistant daily, the information lives in a system with an API, and people ask the same questions repeatedly. Start with one read-only server answering a single question, which takes days and cannot break anything.
Related
It is, when scoped properly. Expose the narrowest set of capabilities that answers the real questions, separate read from write, put approval gates on anything destructive, log every call, and give it its own credentials rather than a shared admin account.
Related
Call, email, or use the contact form. The form asks a few questions up front so the first reply can be useful rather than a request for more information.
Within one business day, usually sooner. If it is urgent, calling is faster than the form.
Related
Alberta, Canada. Work happens remotely with clients across Canada and the United States, over email, video calls and shared preview links.
Related
Regularly. Location has never been a constraint on this kind of work. Being Canadian helps with timezone overlap, invoicing in Canadian dollars and familiarity with the privacy rules, but it is not the deciding factor.
Related
That is the preferred starting point. No pitch deck, no pressure, and no obligation. You will get an honest read on whether the thing is worth building at all.
Related
That is the normal case and completely fine. Describe what is going wrong or what takes too long, and working out the shape of the solution is part of the job.
