The Techstrong.ai stories enterprise IT leaders actually read.
Volume 343, Issue 2 · 14 articles
Phonely's CEO explains why live voice agents favor fast, specialized models and should be measured by value, not cost cuts.
In a Techstrong.ai video, Phonely co-founder and CEO Will Bodewes tells Mike Vizard that callers won't wait for lengthy model reasoning, so voice AI often favors smaller, task-specific models over large general ones.
Open-weight models give control but require scarce expertise, so working with providers that fine-tune models for a task is a middle path.
Voice agents still need data, tools and MCP connections within tight latency limits, plus ongoing training and testing beyond a good demo.
Bodewes urges measuring outcomes like 24/7 coverage and lead conversion rather than call cost alone, and he opposes unsolicited AI cold calls.
Reco's CEO outlines how to discover AI agents, assess their exposure and act, with reduced permissions and human oversight.
Reco co-founder and CEO Ofer Klein tells Mike Vizard that governing AI agents starts with knowing which ones are running and what they can reach, though an inventory alone misses risky combinations of sensitive data access, write permissions and external connections.
His three priorities are to discover agent activity through APIs, browser and endpoint vantage points, assess exposure, and take action.
Excessive permissions can turn one compromised agent into broad access, so trimming rights and watching for behavior changes matter.
Manual review can't keep pace, but AI-assisted defense needs guardrails and human oversight, delivered through a phased plan tied to business needs.
Open-weight models like Alibaba's Qwen give Europe operational control, but not the data, methods and talent sovereignty needs.
Alan Shimel examines Alibaba chairman Joe Tsai's pitch in Turin that open-source AI is Europe's route to technological independence.
Tsai sensibly urged European compute and local adaptation with European industrial data, and running downloadable models on owned infrastructure does bring real operational autonomy.
But open weights are the output of a recipe, not the recipe itself, since they don't include the training data, methods or expertise needed to build the next model.
Shimel argues that Europe must not swap American dependence for Chinese dependence and should invest in domestic compute, research, talent and transparent development.
General-purpose AI can look right while missing the domain detail that matters commercially; expert-labeled data closes the gap.
Tosin Kolawole, building LigaxAI, found that a leading general model could recognize braids but not tell knotless braids from box braids, distinctions that mean different prices and booking times for a salon.
Specialist knowledge is often underrepresented in training data, and labels determine which distinctions a model learns, so expert-verified domain data matters more than volume.
UK figures show AI adoption at 26% of businesses, while only 21% of AI users have integrated tools into their systems.
As AI moves into real workflows, businesses must define what 'good enough' accuracy means for each task's consequences.
Organizations can outsource analysis to AI but not understanding; without traceable evidence, teams risk 'cognitive surrender'.
Stu Sjouwerman notes that nearly nine in ten organizations use AI regularly, according to McKinsey's 2026 State of AI report, including for customer research tasks such as interviews and transcript coding.
But AI is a black box: an AI that pulls three themes from 500 interviews is useful, yet without knowing which conversations and quotes support each conclusion, the result can't be verified.
He warns of 'cognitive surrender', when people accept AI outputs as their own judgment and let speed stand in for confidence.
For pricing, positioning and product decisions, he urges teams to demand traceable evidence behind AI conclusions.
Coding agents follow rules that add steps but ignore rules to stop, so enforcement must live in the harness, not in policy text.
Secure Code Warrior's Pieter Danhieux cites Peking University research showing that coding agents follow rules that add steps, such as disclosure or verification, 77% to 100% of the time when reminded.
Rules requiring them to stop, refuse banned work or hand off to a human saw 0% compliance across all models, because reversing work conflicts with an agent's drive to finish.
His company's AI Trust Index finds repeatable vulnerability patterns by model and framework, and a generic 'follow security best practices' prompt didn't help.
He argues that the AI harness is the real security boundary, with enforced permissions, automated checks, restricted execution and human approval gates.
68% call their AI adoption mature, yet only 28% can quantify ROI, and few have unstructured data ready for AI.
Mike Vizard reports on a Komprise survey of 500 technical decision makers, in which 68% say their AI adoption is maturing or mature but only 28% can quantify ROI, most often in customer service (38%).
Data readiness is the main obstacle, with only 4% having more than 70% of unstructured data ready for lakehouse analysis and only 7% having most data cleaned and classified for AI.
Nearly half struggle to find the right data and clear redundant, obsolete and trivial files, and 46% want to query data in place rather than centralize it.
AI is forcing organizations to address long-standing data management problems.
Measure agentic AI by cost per accepted task, including retries, tool charges and human review, rather than per token.
Burak Oktenli argues that tokens measure consumption, not value, and that the right unit is a completed task that passes acceptance criteria.
To calculate it, divide all attributable costs, including failed, abandoned and escalated attempts, tool fees and human review, by unique accepted results.
In his example, configurations costing $1,000 for 800 accepted tasks and $1,200 for 960 have the same unit cost, but the second delivers more.
He warns that multi-agent delegation can multiply bills through duplicate research and redundant calls even when each step looks cheap, and recommends a cost trace for every production workflow that links usage, financial and review data to outcomes.
A professor's $10M defamation win after years of TikTok lies shows why platform accountability under Section 230 needs reform.
Alan Shimel recounts the case of University of Idaho history chair Rebecca Scofield, who was falsely accused by a TikTok tarot reader, in more than 100 videos, of orchestrating four students' murders.
She sued in December 2022, won a liability ruling in June 2024 and a $10 million damages verdict in February 2026, more than three years later and with no guarantee of payment.
Shimel argues that 'you can always sue' is a poor remedy when platforms control how content spreads but Section 230 limits victims' recourse.
He calls for faster judicial relief and clear platform obligations, with safeguards against abusive takedown requests.
Golden images and IaC assume a defined state, but AI agents work because they do things nobody thought to define.
Andrew Storms draws on his experience building KiloClaw, a hosted OpenClaw deployment, to argue that three decades of predictable infrastructure, from gold images to Terraform and immutable systems, rest on defining acceptable state.
Autonomous agents work precisely because they work out actions and requirements as they go, and customers can't say in advance what they'll need.
That forces an uncomfortable trade-off: tight permissions undercut usefulness, frequent human approvals add friction, and unrestricted execution is dangerous.
He argues that adding more monitoring agents won't solve the trust problem, so infrastructure practices must evolve to control autonomous behavior without removing its flexibility.
A six-layer model of the agent context window, with techniques to cut token waste and block prompt injection.
Nandish Nanjappa treats an agent's context window as its volatile memory and walks through a corporate flight-booking agent to define six layers.
They are system directives and policy anchored at the start, a skill registry of tool schemas, retrieved ground truth, rolling episodic history, working reasoning and execution state.
Optimizations include dynamically loading only the tools relevant to the current sub-goal, prompt caching and distilling tool observations to save tokens.
For security, he recommends wrapping retrieved content in clear structural tags and strictly separating trusted instructions from untrusted data to prevent prompt injection, data exposure and unauthorized tool calls.
Alan Shimel reads Hollywood's darker take on Silicon Valley as a warning that AI firms must not repeat social media's mistakes.
Alan Shimel compares HBO's Silicon Valley, whose founders audiences still rooted for, with the new series Cupertino and Aaron Sorkin's The Social Reckoning about Facebook whistleblower Frances Haugen.
He sees a shift in public perception from celebrating founders to questioning the power and accountability of tech leaders.
Scrutiny of Big Tech has grown while legislation and accountability lag.
As AI companies promise another revolution amid declining public trust, Shimel argues the industry needs to confront its history rather than repeat the mistakes of the social media era.
One-off AI integrations become a maintenance and governance burden; standard access layers like MCP simplify the architecture.
Alonso Robles notes that many organizations connect AI to enterprise systems one connector at a time.
Each integration brings authentication, permissions, API changes and governance work, and that work piles up after the twentieth or thirtieth connector.
He argues enterprise AI needs abstraction: a consistent way for agents to discover trusted business context, through standards such as the Model Context Protocol, without every model needing to understand the underlying infrastructure.
Consistent access patterns also make it easier to enforce permissions, monitor data use and maintain security as models and requirements change.
Sovereign AI has moved beyond data residency paperwork to national autonomy over infrastructure, models and languages.
Russell Fishman argues that sovereignty is no longer just about where data sits and whose law applies, since countries don't want hospitals, power and defense to depend on technology a foreign decision could cut off.
He separates data, operational and technology sovereignty, noting that most companies don't need all three and that conflating them raises cost.
Japan, the Gulf states and the EU, through June legislation on data center capacity and cloud dependence, are funding national capability, including models for underserved languages.
Sovereign platforms are also being built as full stacks rather than raw GPU marketplaces.
More from Techstrong.ai · All IT section newsletters · This week