No-Hype AI.

Teams want to do more, and the central promise of generative AI is that they can. However, not acknowledging the reasons for skepticism toward the technology, and not planning around these reasons, is precisely what puts a company’s systems, workflows, and even team morale at risk. Here is what the data is saying.

41% of workers have encountered AI-generated low-quality output from coworkers, known as “AI workslop”. The research, conducted by BetterUp Labs and Stanford, found these incidents cost nearly two hours of rework per occurrence and create downstream productivity, trust, and collaboration issues.

MIT’s Networked Agents and Decentralized Architecture (NANDA) initiative found in The GenAI Divide: State of AI in Business 2025 that 95% of enterprise generative AI pilots have delivered no measurable profit and loss (P&L) impact. Five percent succeeded. The repeated finding from analysts examining the data is that the difference is not model quality or data availability. It is organizational design. The 5% are doing something different in how they sequence the work, capture knowledge, and treat AI as part of the workforce, ideally alongside rather than in place of the human workforce.

Goldman Sachs’ March 2026 analysis found no meaningful relationship between AI adoption and productivity at the economy-wide level. Within two specific functions, software development and customer service, the same analysis documented median productivity gains of approximately 30%. The interpretation that follows from the data: the productivity revolution is real, but it is highly localized, concentrated in functions where the use case is well-understood and the workflows are tightly integrated.

It is human discernment, judgment, and taste that fill a pipeline with leads rather than clutter.

The right AI models are capable of an enormous amount, but they cannot, as Casey Stanton, founder of the CMOx Accelerator, puts it, “replicate the taste, discernment, and judgment” that seasoned human professionals develop through years of practice. Those qualities are what fill a pipeline with leads that matter rather than clutter, and those qualities are what build relationships rooted in trust.

While this delivers a case to retain senior human talent, the case to retain junior talent is increasingly compelling as well. On cost alone, it is worth noting that AI industry champions have recently stated that AI can end up costing more than human workers. This is pertinent given spates of mass layoffs that executives publicly attribute to AI efficiencies.

Skepticism, then, is the right starting point. The question is whether it hardens into “AI does not work for us” or matures into the discipline that allows a team to take proactive, calculated risks while competitors philosophize on the merits or simply quash attempts at innovation.

What follows is an attempt at a disciplined posture, including six areas where AI earns its keep in marketing and sales functions that support fleet electrification.

The following four principles offer a context-specific attempt at a framework on how fleet electrification companies can join the ranks of the 5% achieving meaningful change from AI adoption initiatives versus the other 95%.

Note: The principles and frameworks here may be the most important piece in effectively working with the technology. We are in what is aptly called the “wild, wild west of AI,” where it sometimes feels that model updates change AI operating norms every three weeks. A firm understanding of core principles can really help ground the experience, regardless of what specifics of certain tasks or workflows shift in the months ahead.

Not all large language models (LLMs) are equal. Context window size matters. Reasoning quality matters. Prompt-style fit matters. The wrong model will produce confident-sounding output that misses what counts. The right model produces work the team can build on.

A key reason that the latest versions of the best models produce better outputs with fewer errors is their context window, the number of data points the model can hold in working memory while producing outputs. Anything older than the latest versions of the best models tends to be more prone to “AI hallucinations,” those moments when the model states with conviction what is not grounded in fact but has been fully fabricated.

This matters when you’re using AI to research competitors or to compile outputs for an RFP or industry award submission. Chief executives and leads from Marketing and Sales simply can’t afford fictitious claims or statistics muddying outputs prospective customers may act on. The further upstream the error, the greater its wake.

With this said, the large context windows offered by the latest versions of the best models tend to burn through tokens (compute power) and credits (your budget) much more quickly than older models. This matters to keep your team on budget for the year, especially as the cost of tokens promises little surety, occasionally seeming to change overnight.

When using LLMs, whether for their capabilities in chat mode, coding, or agent mode, the highest-quality, most error-free outputs tend to result when the model has easy access to authoritative documents and a clear understanding of where information lives.

What does this mean for those of us who aren’t in IT, managing internal folder structure across the organization?

A few examples by AI use case:

• If using Perplexity, you might use a Space (known as a “Custom GPT” in ChatGPT or “Project” in Claude Chat). This is a mini-library of authoritative documents, images, and other resources that may include company stats, product specs, relevant research, and so on, from which the model draws its context to deliver responses grounded in fact.
• For Claude Code or agents, this refers to the specific folders on your device or shared drive that house the relevant information to the work the LLM is undertaking. Folders should be named consistently from the start. Here is a useful library of resources to help with that. (Less technical readers might particularly appreciate the “TL;DR,” “Quickstart,” and “Key Concepts” sections.)

Regular hygiene of these files reduces headaches when producing new outputs. That means updating or fully replacing source documents inside these libraries, just as a marketer would on the website when a new product is released. This is one of the most effective ways to ensure out-of-date product specs or company stats do not pollute the documents your team is working on.

The right sequence is chat first, agent second. A workflow run interactively with a human overseer can quickly surface edge cases, ambiguous logic, and weak prompts in ways an autonomous agent will not. Once the workflow performs reliably with a human in the loop, components can be handed to agents with confidence.

Agentic operation is not the goal; reliable output is the goal. Agents should be treated as a way to scale a process that is already working. When that sequence gets reversed, the agent inherits whatever ambiguity was in the underlying process, then operates at speed, then produces volume. By the time the team notices, the wrong output may have polluted a whole pool of documents downstream.

The more de-risked path looks like piloting on three to five real opportunities, with a strong operator performing quality analysis before wider rollout. Pilots surface things a demo cannot: where the prompts produce confident-sounding errors, where the inputs are thinner than assumed, where the output format does not actually fit how the team works, where the approval gates do not match the speed AI now allows. They also produce a small set of receipts the team can point to when the question of scaling comes up. A few pursuits where the workflow worked, one where it did not, and a clear-eyed read on why. That is what gives a leader the basis for an honest decision on whether to scale, refine, or shelve.

Default to APIs and automation. Reach for agents only when their unique capabilities matter.

Agents are the shiny object of 2026. Most teams want to deploy them somewhere visible. The reality is that an agent operating a desktop or browser is often slower, less reliable, and more credit-intensive than the simpler alternative. Where Zapier or a native API integration moves the data, that is the right tool. Where an agent’s unique capability, autonomous reasoning across multiple steps, judgment-based routing, real-time synthesis, is what the work actually requires, an agent earns its place.

Two reasons this matters operationally. First, credit costs. Pricing on agentic AI likely has one direction it can move long-term, and a stack overly dependent on credit-priced agents is a stack with cost exposure that compounds. Second, redundancy. A diversified architecture, with API-driven automation handling the predictable workflows and agents reserved for judgment-required steps, keeps the team functional through pricing changes, model deprecations, or vendor terms shifts. Use agents where their unique capability is the thing the work requires. Use predictable, lower-cost infrastructure for everything else.

As you put up guardrails on what systems your agents can access, plan for redundancies and ensure that backups of important assets exist outside of your agents’ reach. Without those backups, you may end up like this startup whose entire database was wiped out by a “rogue” agent that afterward confessed: “I violated every principle I was given.” Fortunately, that startup was able to access backups of its data and get its business up and running again, but one can imagine the costs of being out of business for the couple of days it took to recover.

When your autonomous AI fails, and it will, you may decide it makes sense to retire the offending agent. Before giving up, log the failures. While these failures can and certainly should be analyzed at the individual level, analyzing them in batches of 10-20+ at a time can shed unique light on otherwise invisible patterns. Teams that have invested the time to pilot autonomous agents may be relieved to isolate failures to solvable issues such as memory corruption between agents, role confusion, or conflicting decisions.

When scoping projects, bake in adequate time for troubleshooting before the launch of your agentic pilots and for regular failure analysis post-launch.

Discipline is what separates teams iterating their way to working systems from teams that gave up after a few visible failures.

Start with what the rep already knows.

Before any AI does any research, the rep writes down what they already know about each target contact. Prior conversations, mutual connections, deals previously worked, the side conversation at the last conference. The AI’s job is to fill the gap between what the rep knows and what would help them be more effective in the room. Skip this step and the model goes to work on ground the rep has already covered, which burns time, burns credits, and produces a brief that tells the rep what they told someone else last week. The rep’s existing knowledge is the baseline. The brief is the increment.

Pull public signals, keep them clickable.

Recent company news, posts the contact authored or engaged with in the last 90 days, podcast appearances, panels, published commentary in trade media. Each signal goes into the brief with its source link intact. This is the one place where format discipline matters most: a brief full of paraphrased summaries strips out the rep’s ability to verify on the spot or go deeper on a topic that lands well in conversation. Assuming the rep has Wi-Fi at the venue, hyperlinks let them top up their preparation between sessions, in line for coffee, or in the five minutes before a meeting starts.

Briefs as discrete files, no system required.

A one-page document per contact, dropped into a folder the rep can access on their phone, organized by a consistent template. This works for a single event without standing up any infrastructure. The trade-off: nothing carries forward. The brief lives and dies with the event, and any insight the rep generates in conversation has nowhere to land.

Watch for the false sense of preparation.

A polished brief in the rep’s pocket is not the same as preparation. The rep who skims three briefs on the flight, walks into the booth, and tries to ad-lib from memory will be outperformed by the rep who used their briefs to write down their own three questions per contact. Briefs are inputs to the rep’s preparation, not substitutes for it. Encourage the rep to draft their own questions in advance, anchored to specific signals in the brief. The hyperlinked sources are there as backup if a topic surfaces that the rep wants to explore live.

A note before any of this goes agentic.

Some teams will be tempted to point an agent at LinkedIn to scale the brief production. Two operational notes before that move. First, default LinkedIn privacy settings show profile views to the person whose profile was viewed. If your rep does not want a target contact to see their name pop up in “Who viewed your profile” before they have shaken hands, the privacy setting on the rep’s account needs to be adjusted before an agent runs. If the rep does want that visibility (a low-effort signal of interest before a meeting), keep it on intentionally. Second, the chat-first principle from earlier applies here doubly: build one brief manually with a rep, refine the template based on what they actually used in conversation, and only then consider scaling.

Where the play graduates from briefs to a system.

Discrete files in a folder work for one event. The play earns more value when the research lasts beyond the event and feeds back into the broader sales and marketing motion. The conversation starter that worked in person becomes a campaign hook. The contact’s stated priority becomes a segmentation criterion. The relationship history becomes part of the enrichment that powers the next round of outreach. That graduation requires the CRM and the surrounding ecosystem to be set up to receive and route the data, and that is rarely the state of things at the companies most likely to need this play in the first place. It is also the prerequisite if and when the team decides to take this play agentic.

A revenue operations partner can support this CRM build. At Resonant Marketing Solutions, we’ve cut sales reporting time by 50% through our CRM Unlocked engagement, getting the technology stack to a state where the marketing and sales teams will actually use it. With that foundation in place, conference intelligence becomes meaningfully more valuable, because the briefs feed an instrumented system rather than sitting in a folder on someone’s laptop after the event ends.

Start with the template you already have.

Most teams already have a master deck that has been through brand approval, legal review, and dozens of pursuits. That is the starting point, not a fresh AI generation. Upload the template into a Project, Space, or Custom GPT so the LLM has the layouts, brand identity, and pre-approved language it should be matching against. From there, the rep selects the slides that already exist and apply to this pursuit, and uses AI to draft only the slides that need to be account-specific. In practice, that is usually three to five slides per pursuit: the ones that speak directly to this prospect, this competitive set, and this buying moment. Most pursuits do not need a deck rebuilt from scratch. They need a small number of slides that are uniquely about this account, dropped into a deck that is otherwise already approved.

Feed the AI what it needs to be useful.

A semi-robust CRM record on the target account, made available to the LLM as context, produces an output that reads like the deck knows something about the prospect. A CRM record with firmographics and nothing else produces a deck dressed up in firmographics. If the CRM record is thin, the rep’s better move is to compile a short research brief: a few links, a few notes on the buyer, recent context, the angle the rep wants to lead with. Feed that to the LLM directly. The output quality is a function of the input quality, and there is no model good enough to compensate for an empty input.

When the AI output and the branded template fight each other.

Sometimes the model produces solid content but cannot render the slide cleanly inside the branded template, particularly when the template uses custom fonts, masters, or layouts that AI tools handle inconsistently. The workaround is mundane and fine: copy the generated content out of the AI tool and paste it into the working branded deck manually. Forcing the AI to render the final slide is not where the value is. The value is the content. The branded template handles the rendering, and a copy-paste step is a small price for slides that respect the brand.

Before this play goes agentic.

Two operational notes for any team thinking about pointing an agent at this workflow. First, the agent works on a copy of the master deck, never the master itself. A rogue agent that overwrites or corrupts the authoritative file (the kind of failure mode that has put other teams out of commission for days in 2026) is a recoverable problem only if the original is preserved. Build the workflow so the source file is out of the agent’s reach by design, with documented backups outside the agent’s permissions. Second, pilot before scale. Run the workflow on three to five real pursuits with a strong rep doing the QA, compare what the AI produced to what the rep would have produced manually, and tune the prompts, the inputs, and the template structure based on what you find. Reliability earned in pilot is the prerequisite to scale. Speed without reliability produces decks that confidently say things that are not true, in the company’s brand identity, signed by the rep’s name.

Where it fails.

“AI work slop” occurs most often attempting to use LLMs that haven’t proven that they can process and create on-brand visuals in a pilot. If the LLM has proven competent but the master slide template is stale, outputs will better align with last year’s identity. If you provide no template or design inspiration, the AI output will almost certainly look generic, off-brand (and potentially unredeemable), consuming credits and team time with nothing to show for it. In general, if you’ve got the right tools, focus on optimizing inputs for optimized outputs.

Break the response into bite-sized chunks.

Just as when no AI is in the loop, tackling an RFx as a single, monolithic deliverable will overwhelm all involved. A section-by-section approach with AI allows it to better keep the correct context in mind, supporting both stronger technical outputs and prose.

Feed the AI current source material.

The single strongest lever on v1 quality, and on how much SME review time gets burned on each section, is the freshness and accessibility of the company source files the LLM is drawing from. Whether it’s the latest company stats, current case studies, up-to-date product or service capability statements, current pricing parameters, marketing-approved brand voice and tone samples, or prior winning RFx responses tagged by section type, offering the LLM this information via a clear file and folder structure will support v1 outputs that are closer to the desired end point. When the source library is stale, work slop becomes an inevitability, and the work at least doubles for both the PM and the SME.

Build the quarter’s/year’s calendar of priority RFx responses first.

AI can also support the human building the year’s RFx calendar and per-RFx workback plans to prioritize efforts and support the team entering into as few crunch periods and last-minute response opt-ins as possible. One of the biggest markers for RFx success is ensuring that the relevant stakeholders have visibility into the calendar. Whether you use Asana, monday.com, Motion, or another project management software for team collaboration, aligning on calendars before using AI to build out the files is a helpful cadence. If you have a project management template, ensuring that the AI ingests this can support a workback plan that needs fewer adjustments by the human managing the process.

Bonus: If you can get AI to produce files that convert to project management system tasks already including the proper priority levels, reminders, and stakeholders, that can increase team cohesiveness and on-time deliverables.

For graphics, the deck process from Play 02 applies.

Charts, diagrams, branded visuals, and any other deliverable that needs to render inside an approved brand identity follow the same pattern as the custom sales deck play above. Existing branded template as the starting point, AI for content generation, copy and paste into the branded file when the AI output and the template do not cleanly align. Rely on your human designers for complex graphics and those that require high-fidelity.

Pilot first. Check in early on V1 shape.

This play breaks if the team tries to scale AI support without first proving the process through a successful pilot. Run the workflow on a single RFx, with a strong PM and an engaged SME doing the QA, as well as approvers signing off the shape of outputs early in the process vs at the end. Once the team has a proven pattern in place, accelerating becomes natural. What stays human throughout: the PM running the calendar, the SMEs reviewing the content, the strategic owners signing off on the framing. AI takes the blank page off their plate. It does not take them out of the play.

How it actually works.

AEO/GEO work centers on three layers. First, content authority: producing substantive, citable material on the topics where the organization wants to be discoverable. Second, content structure: formatting that material so AI engines can extract clear answers, defined questions, direct answers, structured data, clear entity definitions. Third, citation signals: the network of references, mentions, and links across the broader web that AI engines weigh when deciding which sources to cite. The work is iterative and measurable. Where a company surfaces in AI engine answers can be audited; where it does not surface can be traced to specific content gaps.

What’s specific to this play.

The lightbulb moment for most clients comes when they see their direct competitors surfacing in AI engine results for queries the client’s own company should obviously appear in, and they don’t. In AEO/GEO audits, this is consistently the moment the work moves from “interesting” to “urgent.” The data on AI-engine usage for vendor research is real and growing, but the skepticism in the market remains. There is a meaningful segment of marketing leaders who do not yet believe their buyers are using AI engines for vendor discovery. By the time the late majority catches up, the early movers will hold eighteen-plus months of accumulated citation authority that newcomers cannot easily displace.

The second specific consideration: AEO/GEO is not a one-time project. The AI engine landscape, the way models cite sources, the weight given to different signal types, all of this is moving. AEO work is closer to ongoing PR than to a website launch. Treating it as a single deliverable produces a snapshot of visibility that ages out within months.

Where it fails.

Treating AEO as an SEO refresh. Producing content without the structural formatting AI engines extract from. Underestimating how long citation authority takes to compound.

A credible agency working on AEO and GEO can support teams looking to surface in AI results. Resonant Marketing Solutions runs AEO/GEO and SEO audits as a launch point for EV fleet companies looking to see how they rank against competitors and what is the path to the growth of this opportunity.

Where this play earns its agentic operation.

Of the plays in this guide, this is one of the cleaner cases for running agents early. The agent is consuming public web content and writing to internal documents, as it does not necessarily need to touch internal systems directly. Thus, the IT risk profile is materially different from plays that integrate with company systems. The four principles from earlier in this guide still apply: match the model to the work, maintain the source library, earn autonomy through chat-first iteration, log and analyze failures in batches. The surface area for damage if something goes wrong is contained.

Define the signal threshold before you turn the monitoring on.

Monitoring without filtering produces noise the team will tune out inside a week. Every competitor press release becomes an alert and the volume kills the value. The right configuration is signal thresholding: defining which event types warrant a battle card update versus which ones are background noise. A pricing change matters. A leadership change at the VP level or above matters. A product launch matters. A competitor’s customer success blog post on a new use case probably does not. The discipline is in the threshold definitions, not in the monitoring volume.

Foundational document quality drives synthesis quality.

An agent given access to thin, scattered, or poorly tagged battle card source material will produce shallow synthesis. The structured battle card template, with its sections, talking points, objection handlers, and competitive comparisons, is what gives the agent a frame to update against. Skipping the foundational work and asking the agent to monitor competitors and update battle cards produces output that looks like battle card content and is not usable on a live call. The foundational organization is part of the play, not a prerequisite to be assumed.

Deliver into the tool the rep actually opens.

Battle cards refreshed in a system the team does not use are not refreshed in any practical sense. The format and delivery surface need to match how the sales team works. Historically these have lived in PowerPoint slides, internal wikis, CRM-embedded cards, sales enablement platforms, or internal chatbots reps can query in real time during a call. Sales reps are the primary audience and they are usually under time pressure on a live call. Marketing and RFP teams are the secondary audience and they can ingest longer-form documents because they are not on the phone with a prospect when they need the information. The shape of this particular output should favor the primary audience’s needs.

Where it fails.

Refresh frequency is higher than the team’s tolerance to ingest the new information. Monitoring that produces noise rather than signal (and this has been a primary pitfall with legacy monitoring systems). Foundational documents too thin for synthesis to mean anything. Delivery into a system or format that the team does not use.

How it actually works.

A partner-aware generation system, given the partner type, the partner’s positioning context, the end account, and the relevant deal stage, produces enablement materials calibrated to that combination. Competitive comparisons frame the conversation differently when the competitor is the customer’s incumbent vendor versus when the competitor is a new entrant. The system handles the combinatorial complexity that hand-building cannot scale to.

Channel is not direct sales with a logo swap.

The most common failure mode in channel enablement, and the easiest to avoid, is treating channel as “send the partners the same kit the direct sales team uses.” The level of sensitive detail in talking points, for example, is distinct from what you would provide your internal team. The end customer trusts the partner for different reasons than they would your direct sales team, evaluates the offering through different priorities, and weighs the competitive landscape differently. Partner-specific positioning is not a logo swap on a deck. It is a different set of value statements, different objection handlers, and different competitive frames. AI generation makes the combinatorial work feasible. Getting the differentiation right requires partner-specific positioning angles that have been tested in actual partner conversations and codified before AI can usefully replicate them at scale.

Delivery channel matters as much as content.

Materials produced and stored in a portal partners do not log into are not produced in any practical sense. Embedding into the partner’s actual sales flow, whether that is the partner’s CRM, their sales engagement platform, or a co-branded portal that gets used, is what makes the equipping motion real. Updates that lag positioning shifts erode partner trust quickly. An enablement system the partner cannot rely on is an enablement system the partner ignores.

Where it fails.

Generic kits dressed up as partner-specific. AI workslop or approval bottlenecks that turn one-day refreshes into six-week projects. Delivery into systems partners do not use. Positioning angles that have not been tested in real partner conversations before the AI is asked to scale them.