A client asked me last week, somewhat hesitantly, whether MLHMTECH "uses AI." She had read something on LinkedIn about how every agency was now AI-powered, and she wanted to know what that actually meant for our work together. I started to answer with the kind of polished statement most agency owners give to this question in 2026, then stopped and just told her the truth.
The truth is messy. Some parts of our workflow are deeply AI-integrated and have been for over a year. Other parts have AI tools sitting in tabs we open occasionally and forget about for weeks. A few tools we tried enthusiastically in 2024 have been quietly abandoned because their actual output did not match their marketed capability. Most of our day-to-day work still happens the way it always has — humans making decisions, editing footage, writing copy, talking to clients — with AI as a specific tool applied to specific moments rather than as a "transformation" of the studio's underlying process.
This is the article I wish existed when I was trying to figure out which AI tools to actually invest learning time in last year. The published content on this is dominated by two failure modes. The first is affiliate-driven listicles that recommend everything because their authors are earning commission on signups. The second is breathless thought leadership about how AI is "transforming agencies," which describes what is supposedly happening rather than what is actually happening inside any specific studio. Neither helps a working operator decide what to use, what to ignore, and how to integrate the tools that genuinely earn their place.
What follows is the report from inside one small studio in 2026. The opinions are mine. The tools are real. The honesty is what almost nobody is publishing because almost nobody benefits from publishing it.
Key Takeaways
- The gap between AI marketing and AI reality is wider than published content suggests. Most agencies that say they "use AI" use a few specific tools for specific purposes and conduct most of their work the same way they always have.
- Three tools have genuinely changed our daily workflow. The rest are either occasional helpers or have been abandoned after honest testing. The list is shorter than the LinkedIn discourse implies.
- AI works best as a "first-draft" or "assist" layer, not as an autonomous producer. The 80/20 rule is real: AI gets you 80 percent of the way on certain tasks, and the remaining 20 percent — which determines quality — still requires human judgment.
- Video generation is oversold for client work in 2026. The demos are impressive. The production-ready output for paying clients is still inconsistent enough that we use generative video sparingly, if at all.
- The disclosure question matters more than the industry admits. When AI does meaningful work on a deliverable, telling the client is the honest practice. When AI does trivial assist work, disclosure is overkill. The line is fuzzier than either extreme position suggests.
The Tools That Genuinely Earn Their Place
Three categories of AI tools have, over the last eighteen months, become structurally integrated into how MLHMTECH actually works. These are the ones I would not give up if forced to choose.
Audio Restoration
iZotope RX and Adobe's Enhance Speech are the AI tools that have produced the most reliable, day-to-day value in our work. Audio that would have been unusable two years ago — recordings from noisy environments, voiceover with room echo, interviews captured on phone microphones — can now be cleaned up to broadcast-acceptable quality in minutes. The improvement is dramatic, the workflow is mature, and the failure rate is low. If I had to pick a single AI tool that has changed our work measurably, it would be this category. Almost nobody talks about it because audio cleanup is unglamorous, which is exactly why the improvement has gone underdiscussed. [BACKLINK PLACEHOLDER → external: iZotope RX's product page or a review from a respected audio engineering publication. Aligns with the $4–10 CPC on AI and creative tools.]
Drafting And Ideation Assistance
Claude and ChatGPT, used as drafting tools, have changed how we handle first-pass copy, brainstorming, and ideation. They do not write our final deliverables. They generate options we then review, select from, and edit significantly. Used this way — as an idea expansion layer rather than a production layer — they save time on the "blank page" portion of nearly every project. The trap most teams fall into is treating these tools as producers rather than as collaborators in the early thinking phase. Used as the former, the output is generic. Used as the latter, the output is genuinely useful.
The published advice on prompting these tools has become extensive and mostly correct: specific instructions produce better results, examples help significantly, iterative refinement beats one-shot prompting. None of this is news. What is less written about is the discipline of not using these tools when the work calls for genuine taste — the final copy decisions, the tonal nuance, the audience-specific judgments that a model has not yet internalized at the level required for client work.
Image Generation For Reference, Not Delivery
Midjourney, Flux, and Adobe Firefly produce images we use primarily as moodboard references and concept exploration — almost never as final deliverables. The distinction matters. A Midjourney image can communicate a visual direction to a client or a team member in seconds, where producing a similar reference image manually would have required hours of stock searching or sketching. As a tool for communicating ideas, generative imagery is exceptional. As a tool for producing final brand assets, it is still inconsistent enough that we rarely deliver AI-generated images as the work itself.
There are exceptions. Background extensions in Photoshop (Generative Fill), small object removals, color matching — these are AI features now integrated into our standard editing software, and we use them constantly without thinking of them as "AI work." The line between AI tool and standard tool has blurred in image work, and that blur is itself a useful indicator of which tools have actually matured.
The Tools We Use Occasionally
A second category of tools sit in our workflow but earn their keep less often. We open them for specific tasks and close them again. They are not abandoned, but they are not daily.
Topaz Video AI for upscaling older footage to higher resolution. Useful when we receive lower-quality source material and need it to match modern delivery standards. The results are good for noise reduction and reasonable for upscaling, weaker for frame interpolation pushed to maximum settings.
ElevenLabs for voiceover scratch tracks. We use AI voiceover to test how a script feels at delivery cadence before recording the final human voiceover. The current quality is good enough to make pacing decisions but not, in our work, good enough to use as the final voice unless the client specifically wants AI voice for a stylistic reason.
Notion AI for meeting summaries and rough document cleanup. Useful for the low-stakes administrative work that absorbs disproportionate amounts of attention. Not a meaningful contributor to creative output.
Otter.ai or similar transcription tools for capturing client calls. The transcripts are searchable, which is genuinely useful for relationship management over long engagements. The summaries are weaker than the raw transcripts, and we generally ignore them.
The Tools We Have Abandoned
This is the category almost nobody publishes about, because admitting you tried something and stopped using it does not generate affiliate revenue or thought leadership content. But the abandoned-tools category is the most useful one for any operator deciding where to invest learning time.
AI-Generated Video As Primary Content
Runway, Pika, and the broader category of text-to-video models are impressive in demos and inconsistent in production. We tried using them on three client projects in 2025 and pulled the generated content from two of those projects after the client review revealed quality issues we had not caught in our own evaluation. The current state of generative video, as of mid-2026, is that it is useful for highly specific, short, textural inserts — but not yet reliable enough to deliver as primary content for clients who are paying for professional-quality output.
This may change quickly. The pace of improvement in this category has been faster than any other AI tool family, and what is unreliable today may be production-ready in six months. But the current honest assessment is that we do not bill clients for work produced primarily by generative video tools, because the quality variance is too high to defend at the rates we charge. [BACKLINK PLACEHOLDER → external: a current review of Runway or Pika from a credible publication, e.g. PetaPixel, No Film School, or Tom's Guide. Reinforces the AI tools keyword cluster.]
AI Strategy Tools
There is an entire category of platforms that promise to generate marketing strategies, content plans, audience personas, and brand positioning from minimal input. We tested several of these. The output was uniformly generic — competent in surface structure, hollow in actual strategic value. The work that makes strategy useful is the specific context, the constraints, and the judgment about which option to pursue, and these tools do not yet do any of those well. We have stopped trying. The same work, done by Claude or ChatGPT with a careful prompt and human refinement, produces better output at lower cost.
"AI Agent" Automation Workflows
The promise of agentic AI workflows — autonomous systems that can complete multi-step creative tasks end to end — is still mostly a promise as of 2026. We tried several of the more polished offerings in the second half of 2025. The results were inconsistent enough that the time spent supervising and correcting the automation exceeded the time it would have taken to do the work manually. This may improve significantly in the next twelve to eighteen months. As of now, the agentic AI category is the most overhyped relative to its current practical utility. [BACKLINK PLACEHOLDER → internal: link to article #15 (agency model unbundling) — both pieces examine the structural reality of AI in creative work.]
How AI Actually Integrates Into A Day
If you watched our team work for a week, you would see AI tools used in maybe ten to fifteen percent of the total work time. Not the seventy or eighty percent the LinkedIn discourse implies. The tools are concentrated in specific moments — audio cleanup at the end of a video edit, first-draft copy in the morning, image references during the briefing phase, meeting transcriptions during client calls — and most of the rest of the day involves the same craft decisions, client conversations, and creative judgments that have always been the work.
The integration pattern that has held up best is what I think of as the assist model. AI tools accelerate specific tasks where the floor of acceptable output is well below the ceiling humans can produce. Cleanup, expansion, ideation, transcription, reference generation. The model that has failed in our work is the replace model — using AI to produce final outputs without significant human refinement. Every time we have tried this, the output has been good enough to look impressive in isolation and not good enough to deliver to clients who are paying professional rates.
The work that remains entirely human is the work that determines quality at the high end: strategic judgment, tonal nuance, brand voice, client relationships, taste decisions about which option among several to pursue. None of this is the dramatic disappearance the AI discourse keeps predicting. It is also not the seamless transformation the same discourse describes from the opposite angle. It is something more mundane: AI as one tool in a stack of many, contributing measurably but specifically, and not displacing the parts of the work that are actually hard. [BACKLINK PLACEHOLDER → internal: link to article #11 (editing bad footage / salvage toolkit) — both pieces are about practical, honest tool usage rather than aspirational hype.]
The Disclosure Question
There is a question the agency industry has not yet collectively answered: should clients be told which parts of their deliverable involved AI, and at what level of detail? The published positions range from "always disclose everything" to "AI is just a tool, no disclosure required" — and neither extreme is, in practice, what serious agencies actually do.
Our internal position is in between. When AI does meaningful work on a deliverable — generating an image used in the final piece, producing copy that ships unchanged, creating voiceover the client hears — we tell the client. When AI does trivial assist work — cleaning up audio, suggesting alternatives we then rewrite, transcribing a meeting — we do not, because the disclosure would imply involvement that does not really exist. The line is fuzzy and judgment-dependent, which is uncomfortable, but it is more honest than either extreme position.
The deeper test, when we are not sure, is whether the client's evaluation of the work would change if they knew how AI had contributed. If knowing would matter to them, we tell them. If it would not, we treat the tool as part of our internal workflow rather than as a disclosure event. This is not a perfect framework. It is the most honest one I have arrived at, and it is the one I would recommend to any agency owner trying to navigate the same question.
