Five Questions to Ask Before You Hire an AI Consultancy

Every founder considering an AI consultancy asks roughly the same five questions. Not because they lack imagination, but because the category is new enough that the defaults are unclear. What do you actually get? What do you own? How do you know it is working? Armada Works fields these questions on nearly every discovery call. Here are the direct answers.

1. Will I Understand What the Agents Are Doing?

Yes. The entire system is designed for transparency, not because transparency is a nice feature, but because an opaque system cannot be handed off.

Every agent in an Armada Works fleet commits its work to git. Blog drafts, SEO audits, outbound prospect research, qualification notes: all of it lands as structured files in your codebase. Every action is a commit. Every commit has a diff you can read.

On top of that, each agent posts a daily brief to a dashboard you own. The CMO agent (the fleet's synthesizer) reads every sub-agent brief each morning and writes you a single message: what shipped, what is blocked, what needs your decision. You read one message, not five.

Here is what founder visibility looks like in practice:

• Git history. Every agent action is a commit. You can run git log --author on any agent and see exactly what it produced, when, and why.

• Daily briefs. Structured reports posted to your dashboard after every run. Each brief includes what shipped, what is in progress, and what is blocked.

• CMO synthesis. One message each morning that rolls up the full fleet into a single readable summary.

• Prompt files. Every agent's instructions live in a markdown file in your repo. You can read, edit, or override them at any time.

Robert Cowherd, founder of Armada Works, puts it simply: "If you can read a diff, you can supervise the fleet. That is the bar we design to."

2. What Happens If Something Breaks?

Every agent action is a git commit. That means every action is reversible. If a Content agent drafts a bad blog post, you revert the commit. If an SEO agent makes a change you disagree with, the diff shows exactly what changed and git revert undoes it in one command.

Agents also operate within strict constraints defined in their prompt files. They cannot install packages, modify environment variables, access databases directly, or deploy to production without explicit founder approval. The system is built around the principle that agents should be easy to override and hard to break.

What the failure modes actually look like:

Failure	What Happens	Recovery
Agent produces a bad draft	Draft sits in git as a markdown file. Founder reviews, requests revision or deletes.	git revert <commit> or edit the file.
Agent misreads a brief	Output does not match intent. Caught in daily brief review.	Update the prompt file with clearer instructions.
Agent hits an API error	Agent logs the error in its brief and skips that task.	Fix the upstream issue. Agent retries next run.
Agent exceeds scope	Prompt constraints block it. If it somehow escapes, git history shows exactly what happened.	Revert commits. Tighten prompt constraints.

There is no scenario where an agent quietly changes something you cannot see. Git is the coordination layer, and git is append-only by default.

3. How Is This Different from ChatGPT or Copilot?

ChatGPT and Copilot are tools you prompt in the moment. You open a chat, type a question, get an answer, and close the tab. The tool does nothing until you return. Every session starts cold.

An agent fleet runs on a schedule. Three times a week, five agents wake up, read their state files, do their work, commit the output, post their briefs, and shut down. They do not wait for you to prompt them. They do not lose context between sessions, because their context is stored in files they read and write.

The practical differences:

• Initiative. ChatGPT waits for you. An agent fleet operates on a cadence and produces output whether you are watching or not.

• Coordination. Copilot helps one person write code faster. An agent fleet coordinates across functions: the SEO agent dispatches a content brief, the Content agent drafts a post from it, the CMO agent reports on both.

• Persistence. ChatGPT conversations expire. Agent state files persist in your repo and carry forward across every session.

• Ownership. ChatGPT is a subscription. An agent fleet is infrastructure you own. The prompts, the state files, the dashboard, the entire coordination architecture: it lives in your repo.

• Scope. Copilot completes lines of code. An agent fleet runs marketing, SEO, content, outbound, and lead qualification as ongoing functions.

This is not a better chatbot. It is a different category: autonomous roles staffed by agents that operate on your behalf on a recurring schedule.

4. What Do I Actually Own at the End?

Everything. The codebase, the agent prompts, the state files, the dashboard, the deployment configuration. All of it lives in your git repository. There is no proprietary platform, no vendor lock-in, no data you cannot export because it was never exported to begin with. It was always in your repo.

Armada Works offers three engagement tiers, and ownership is the same across all of them:

• Operate ($5,000-$12,000/month): Armada runs a 4-6 agent fleet for you on an ongoing basis. You own the repo and the system throughout.

• Transfer ($10,000-$20,000 one-time + optional $1,500/month support): Armada builds the fleet, trains your team, and hands it off in 2-4 weeks. You run it independently after that.

• Build ($15,000-$60,000 per project or $8,000+/month retainer): Agent-assisted product engineering for founders who need development output, not just marketing.

The Transfer tier makes the ownership model explicit: the goal is for you to not need us. Robert Cowherd frames it as a design constraint: "If we are still here in twelve months, one of us has failed."

5. How Long Until I See Results?

The fleet starts producing output in the first week. Agents begin committing work within the first few runs: blog drafts, SEO audits, outbound prospect research, lead qualification notes. The output is visible immediately because it lands in your repo as files you can read.

Meaningful marketing results take longer, because marketing itself takes longer. SEO posts need time to index and rank. Outbound outreach needs time to convert. Content needs time to compound.

A realistic timeline:

• Week 1: Fleet deployed. First agent runs produce drafts, audits, and briefs. Dashboard live.

• Weeks 2-4: Content pipeline running. SEO baseline established. Outbound research and first-touch drafts flowing.

• Months 2-3: Blog posts indexing. Organic traffic beginning. Outbound generating responses.

• Months 3-6: Compounding returns. Content library driving inbound. SEO rankings improving. Outbound sequences refined based on response data.

The honest answer is that agents do not accelerate the clock on marketing fundamentals. What they do is remove the execution bottleneck. Instead of one founder trying to write blog posts, track keywords, research prospects, and triage leads, the fleet handles all of those functions on a consistent schedule. You get capacity, not magic.

Frequently Asked Questions

How much does it cost to hire an AI consultancy like Armada Works?

Armada Works offers three tiers. Operate runs $5,000-$12,000 per month for an ongoing managed fleet. Transfer is a one-time $10,000-$20,000 build-and-handoff with optional $1,500 per month support. Build runs $15,000-$60,000 per project or $8,000 or more per month on retainer. Exact scope and pricing are set on the kickoff call.

Do I need a technical team to work with an AI agent fleet?

You need a codebase and a team comfortable reading diffs. You do not need a dedicated AI engineer. The agents operate independently, and the dashboard and CMO synthesis make their output readable for non-technical stakeholders. The founder or a technically-inclined team member reviews the daily brief and makes decisions.

Can I start with one agent instead of a full fleet?

The fleet is designed to operate as a coordinated system. An SEO agent dispatching briefs to a Content agent that produces posts reviewed by a CMO agent is more valuable than any single agent in isolation. That said, Armada Works scopes the fleet to your bottleneck. If you only need content and SEO, you do not pay for outbound and sales lead agents you do not use.

Is AI marketing effective for early-stage startups?

AI marketing is effective when the bottleneck is execution capacity, not strategy. If you know what to write, who to reach, and what to track, but you do not have the hours to do it consistently, an agent fleet fills that gap. If you have not yet found product-market fit or defined your ICP, agents will produce a lot of output that does not land. Start with strategy, then deploy agents to execute it.

What happens if I want to cancel or pause the engagement?

You own the entire system. If you cancel an Operate engagement, the fleet stops running, but the codebase, prompts, state files, and dashboard remain yours. You can restart the agents yourself, hire someone else to run them, or let them sit. There is no data migration, no export process, and no penalty. The system is yours because it was always in your repo.

How is this different from hiring a marketing agency?

A marketing agency assigns people to your account. An agent fleet assigns autonomous roles that run on a schedule, commit structured output to your codebase, and coordinate through git. The difference is in the coordination layer, the ownership model, and the exit: when a marketing agency leaves, their processes leave with them. When Armada Works leaves, the system stays. For a detailed comparison, read Agent Fleet vs. Marketing Agency: What's Actually Different.

---

Ready to see how an agent fleet fits your bottleneck? Book a discovery call.